Mercurial > mplayer.hg
annotate postproc/postprocess.c @ 2158:508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
author | arpi |
---|---|
date | Wed, 10 Oct 2001 22:13:27 +0000 |
parents | |
children | 795f3d022657 |
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 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
20 C MMX MMX2 |
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 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
23 doVertLowPass E e |
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 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
26 isHorizMinMaxOk a |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
27 doHorizLowPass E a |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
28 doHorizDefFilter E a |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
29 deRing |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
30 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
31 E = Exact implementation |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
32 e = allmost exact implementation |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
33 a = alternative / approximate impl |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
34 c = checked against the other implementations (-vo md5) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
35 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
36 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
37 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
38 TODO: |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
39 verify that everything workes as it should |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
40 reduce the time wasted on the mem transfer |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
41 implement dering |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
42 implement everything in C at least |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
43 figure range of QP out (assuming <256 for now) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
44 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
|
45 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
|
46 move YScale thing to the end instead of fixing QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
47 ... |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
48 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
49 Notes: |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
50 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
51 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
52 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
53 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
54 #include <inttypes.h> |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
55 #include <stdio.h> |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
56 #include "../config.h" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
57 #include "postprocess.h" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
58 //#undef HAVE_MMX2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
59 //#undef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
60 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
61 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
62 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
63 static uint64_t packedYOffset= 0x0000000000000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
64 static uint64_t packedYScale= 0x0100010001000100LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
65 static uint64_t w05= 0x0005000500050005LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
66 static uint64_t w20= 0x0020002000200020LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
67 static uint64_t w1400= 0x1400140014001400LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
68 static uint64_t bm00000001= 0x00000000000000FFLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
69 static uint64_t bm00010000= 0x000000FF00000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
70 static uint64_t bm00001000= 0x00000000FF000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
71 static uint64_t bm10000000= 0xFF00000000000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
72 static uint64_t bm10000001= 0xFF000000000000FFLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
73 static uint64_t bm11000011= 0xFFFF00000000FFFFLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
74 static uint64_t bm00011000= 0x000000FFFF000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
75 static uint64_t bm00110011= 0x0000FFFF0000FFFFLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
76 static uint64_t bm11001100= 0xFFFF0000FFFF0000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
77 static uint64_t b00= 0x0000000000000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
78 static uint64_t b02= 0x0202020202020202LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
79 static uint64_t b0F= 0x0F0F0F0F0F0F0F0FLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
80 static uint64_t bFF= 0xFFFFFFFFFFFFFFFFLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
81 static uint64_t b7E= 0x7E7E7E7E7E7E7E7ELL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
82 static uint64_t b7C= 0x7C7C7C7C7C7C7C7CLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
83 static uint64_t b3F= 0x3F3F3F3F3F3F3F3FLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
84 static uint64_t temp0=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
85 static uint64_t temp1=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
86 static uint64_t temp2=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
87 static uint64_t temp3=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
88 static uint64_t temp4=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
89 static uint64_t temp5=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
90 static uint64_t pQPb=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
91 static uint8_t tempBlock[16*16]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
92 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
93 int hFlatnessThreshold= 56 - 16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
94 int vFlatnessThreshold= 56 - 16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
95 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
96 //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
|
97 double maxClippedThreshold= 0.01; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
98 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
99 int maxAllowedY=255; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
100 //FIXME can never make a movieŽs black brighter (anyone needs that?) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
101 int minAllowedY=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
102 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
103 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
104 static inline long long rdtsc() |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
105 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
106 long long l; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
107 asm volatile( "rdtsc\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
108 : "=A" (l) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
109 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
110 // printf("%d\n", int(l/1000)); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
111 return l; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
112 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
113 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
114 static inline void prefetchnta(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
115 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
116 asm volatile( "prefetchnta (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
117 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
118 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
119 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
120 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
121 static inline void prefetcht0(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
122 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
123 asm volatile( "prefetcht0 (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
124 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
125 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
126 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
127 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
128 static inline void prefetcht1(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
129 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
130 asm volatile( "prefetcht1 (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
131 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
132 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
133 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
134 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
135 static inline void prefetcht2(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
136 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
137 asm volatile( "prefetcht2 (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
138 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
139 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
140 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
141 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
142 //FIXME? |255-0| = 1 (shouldnt be a problem ...) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
143 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
144 * Check if the middle 8x8 Block in the given 8x10 block is flat |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
145 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
146 static inline bool isVertDC(uint8_t src[], int stride){ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
147 // return true; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
148 int numEq= 0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
149 src+= stride; // src points to begin of the 8x8 Block |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
150 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
151 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
152 // "int $3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
153 "pushl %1\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
154 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
155 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
156 "movq (%1), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
157 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
158 "movq (%1), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
159 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
160 "paddb %%mm7, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
161 "pcmpgtb %%mm6, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
162 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
163 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
164 "movq (%1), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
165 "psubb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
166 "paddb %%mm7, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
167 "pcmpgtb %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
168 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
169 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
170 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
171 "movq (%1), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
172 "psubb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
173 "paddb %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
174 "pcmpgtb %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
175 "paddb %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
176 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
177 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
178 "movq (%1), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
179 "psubb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
180 "paddb %%mm7, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
181 "pcmpgtb %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
182 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
183 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
184 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
185 "movq (%1), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
186 "psubb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
187 "paddb %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
188 "pcmpgtb %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
189 "paddb %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
190 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
191 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
192 "movq (%1), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
193 "psubb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
194 "paddb %%mm7, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
195 "pcmpgtb %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
196 "paddb %%mm1, %%mm0 \n\t" |
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 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
199 "movq (%1), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
200 "psubb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
201 "paddb %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
202 "pcmpgtb %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
203 "paddb %%mm2, %%mm0 \n\t" |
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 " \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
206 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
207 "psrlw $8, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
208 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
209 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
210 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
211 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
212 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
213 "psrlq $32, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
214 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
215 "popl %1\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
216 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
217 : "=r" (numEq) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
218 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
219 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
220 // printf("%d\n", numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
221 numEq= (256 - (numEq & 0xFF)) &0xFF; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
222 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
223 // int asmEq= numEq; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
224 // numEq=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
225 // uint8_t *temp= src; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
226 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
227 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
228 for(int y=0; y<BLOCK_SIZE-1; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
229 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
230 if(((src[0] - src[0+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
231 if(((src[1] - src[1+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
232 if(((src[2] - src[2+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
233 if(((src[3] - src[3+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
234 if(((src[4] - src[4+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
235 if(((src[5] - src[5+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
236 if(((src[6] - src[6+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
237 if(((src[7] - src[7+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
238 src+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
239 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
240 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
241 /* if(abs(numEq - asmEq) > 0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
242 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
243 printf("\nasm:%d c:%d\n", asmEq, numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
244 for(int y=0; y<8; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
245 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
246 for(int x=0; x<8; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
247 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
248 printf("%d ", temp[x + y*stride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
249 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
250 printf("\n"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
251 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
252 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
253 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
254 return numEq > vFlatnessThreshold; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
255 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
256 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
257 static inline bool isVertMinMaxOk(uint8_t src[], int stride, int QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
258 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
259 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
260 int isOk; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
261 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
262 // "int $3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
263 "movq (%1, %2), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
264 "movq (%1, %2, 8), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
265 "movq %%mm0, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
266 "psubusb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
267 "psubusb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
268 "por %%mm1, %%mm0 \n\t" // ABS Diff |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
269 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
270 "movq pQPb, %%mm7 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
271 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
272 "psubusb %%mm7, %%mm0 \n\t" // Diff <= 2QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
273 "pcmpeqd b00, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
274 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
275 "pcmpeqd bFF, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
276 // "movd %%mm0, (%1, %2, 4)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
277 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
278 : "=r" (isOk) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
279 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
280 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
281 return isOk; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
282 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
283 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
284 int isOk2= true; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
285 for(int x=0; x<BLOCK_SIZE; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
286 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
287 if(abs((int)src[x + stride] - (int)src[x + (stride<<3)]) > 2*QP) isOk2=false; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
288 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
289 /* if(isOk && !isOk2 || !isOk && isOk2) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
290 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
291 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
|
292 for(int y=0; y<9; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
293 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
294 for(int x=0; x<8; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
295 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
296 printf("%d ", src[x + y*stride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
297 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
298 printf("\n"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
299 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
300 } */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
301 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
302 return isOk2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
303 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
304 |
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 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
307 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
308 * Do a vertical low pass filter on the 8x10 block (only write to the 8x8 block in the middle) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
309 * useing the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
310 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
311 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
|
312 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
313 // QP= 64; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
314 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
315 #ifdef HAVE_MMX2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
316 asm volatile( //"movv %0 %1 %2\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
317 "pushl %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
318 "movq pQPb, %%mm0 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
319 // "movq bFF , %%mm0 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
320 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
321 "movq (%0), %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
322 "movq (%0, %1), %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
323 "movq %%mm5, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
324 "movq %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
325 "psubusb %%mm6, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
326 "psubusb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
327 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
328 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
329 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF |
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 "pand %%mm2, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
332 "pandn %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
333 "por %%mm2, %%mm6 \n\t"// First Line to Filter |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
334 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
335 "movq (%0, %1, 8), %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
336 "leal (%0, %1, 4), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
337 "leal (%0, %1, 8), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
338 "subl %1, %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
339 "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
|
340 "movq (%0, %1, 8), %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
341 "movq %%mm5, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
342 "movq %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
343 "psubusb %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
344 "psubusb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
345 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
346 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
347 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
348 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
349 "pand %%mm2, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
350 "pandn %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
351 "por %%mm2, %%mm7 \n\t" // First Line to Filter |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
352 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
353 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
354 // 1 2 3 4 5 6 7 8 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
355 // %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
|
356 // 6 4 2 2 1 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
357 // 6 4 4 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
358 // 6 8 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
359 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
360 "movq %%mm6, %%mm2 \n\t" //1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
361 "movq %%mm6, %%mm3 \n\t" //1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
362 "paddusb b02, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
363 "psrlw $2, %%mm3 \n\t" //1 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
364 "pand b3F, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
365 "psubb %%mm3, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
366 "movq (%0, %1), %%mm0 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
367 "movq %%mm0, %%mm1 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
368 "paddusb b02, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
369 "psrlw $2, %%mm0 \n\t" // 1 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
370 "pand b3F, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
371 "paddusb %%mm2, %%mm0 \n\t" //3 1 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
372 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
373 "movq (%0, %1), %%mm0 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
374 "movq %%mm0, %%mm1 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
375 "pavgb %%mm6, %%mm0 \n\t" //1 1 /2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
376 "pavgb %%mm6, %%mm0 \n\t" //3 1 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
377 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
378 "movq (%0, %1, 4), %%mm2 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
379 "movq %%mm2, %%mm5 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
380 "pavgb (%%eax), %%mm2 \n\t" // 11 /2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
381 "pavgb (%0, %1, 2), %%mm2 \n\t" // 211 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
382 "movq %%mm2, %%mm3 \n\t" // 211 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
383 "movq (%0), %%mm4 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
384 "pavgb %%mm4, %%mm3 \n\t" // 4 211 /8 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
385 "pavgb %%mm0, %%mm3 \n\t" //642211 /16 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
386 "movq %%mm3, (%0) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
387 // 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
|
388 "movq %%mm1, %%mm0 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
389 "pavgb %%mm6, %%mm0 \n\t" //1 1 /2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
390 "movq %%mm4, %%mm3 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
391 "pavgb (%0,%1,2), %%mm3 \n\t" // 1 1 /2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
392 "pavgb (%%eax,%1,2), %%mm5 \n\t" // 11 /2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
393 "pavgb (%%eax), %%mm5 \n\t" // 211 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
394 "pavgb %%mm5, %%mm3 \n\t" // 2 2211 /8 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
395 "pavgb %%mm0, %%mm3 \n\t" //4242211 /16 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
396 "movq %%mm3, (%0,%1) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
397 // mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
398 "pavgb %%mm4, %%mm6 \n\t" //11 /2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
399 "movq (%%ebx), %%mm0 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
400 "pavgb (%%eax, %1, 2), %%mm0 \n\t" // 11/2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
401 "movq %%mm0, %%mm3 \n\t" // 11/2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
402 "pavgb %%mm1, %%mm0 \n\t" // 2 11/4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
403 "pavgb %%mm6, %%mm0 \n\t" //222 11/8 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
404 "pavgb %%mm2, %%mm0 \n\t" //22242211/16 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
405 "movq (%0, %1, 2), %%mm2 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
406 "movq %%mm0, (%0, %1, 2) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
407 // 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
|
408 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
409 "pavgb (%%ebx), %%mm0 \n\t" // 11 /2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
410 "pavgb %%mm0, %%mm6 \n\t" //11 11 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
411 "pavgb %%mm1, %%mm4 \n\t" // 11 /2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
412 "pavgb %%mm2, %%mm1 \n\t" // 11 /2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
413 "pavgb %%mm1, %%mm6 \n\t" //1122 11 /8 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
414 "pavgb %%mm5, %%mm6 \n\t" //112242211 /16 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
415 "movq (%%eax), %%mm5 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
416 "movq %%mm6, (%%eax) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
417 // 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
|
418 "movq (%%eax, %1, 4), %%mm6 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
419 "pavgb %%mm7, %%mm6 \n\t" // 11 /2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
420 "pavgb %%mm4, %%mm6 \n\t" // 11 11 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
421 "pavgb %%mm3, %%mm6 \n\t" // 11 2211 /8 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
422 "pavgb %%mm5, %%mm2 \n\t" // 11 /2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
423 "movq (%0, %1, 4), %%mm4 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
424 "pavgb %%mm4, %%mm2 \n\t" // 112 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
425 "pavgb %%mm2, %%mm6 \n\t" // 112242211 /16 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
426 "movq %%mm6, (%0, %1, 4) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
427 // mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
428 "pavgb %%mm7, %%mm1 \n\t" // 11 2 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
429 "pavgb %%mm4, %%mm5 \n\t" // 11 /2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
430 "pavgb %%mm5, %%mm0 \n\t" // 11 11 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
431 "movq (%%eax, %1, 2), %%mm6 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
432 "pavgb %%mm6, %%mm1 \n\t" // 11 4 2 /8 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
433 "pavgb %%mm0, %%mm1 \n\t" // 11224222 /16 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
434 // "pxor %%mm1, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
435 "movq %%mm1, (%%eax, %1, 2) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
436 // mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
437 "pavgb (%%ebx), %%mm2 \n\t" // 112 4 /8 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
438 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
439 "pavgb %%mm0, %%mm6 \n\t" // 1 1 /2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
440 "pavgb %%mm7, %%mm6 \n\t" // 1 12 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
441 "pavgb %%mm2, %%mm6 \n\t" // 1122424 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
442 // "pxor %%mm6, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
443 "movq %%mm6, (%%ebx) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
444 // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
445 "pavgb %%mm7, %%mm5 \n\t" // 11 2 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
446 "pavgb %%mm7, %%mm5 \n\t" // 11 6 /8 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
447 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
448 "pavgb %%mm3, %%mm0 \n\t" // 112 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
449 "pavgb %%mm0, %%mm5 \n\t" // 112246 /16 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
450 // "pxor %%mm5, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
451 // "movq pQPb, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
452 "movq %%mm5, (%%eax, %1, 4) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
453 "popl %0\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
454 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
455 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
456 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
457 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
458 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
459 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
460 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
461 const int l1= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
462 const int l2= stride + l1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
463 const int l3= stride + l2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
464 const int l4= stride + l3; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
465 const int l5= stride + l4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
466 const int l6= stride + l5; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
467 const int l7= stride + l6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
468 const int l8= stride + l7; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
469 const int l9= stride + l8; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
470 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
471 for(int x=0; x<BLOCK_SIZE; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
472 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
473 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
|
474 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
|
475 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
476 int sums[9]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
477 sums[0] = first + src[l1]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
478 sums[1] = src[l1] + src[l2]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
479 sums[2] = src[l2] + src[l3]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
480 sums[3] = src[l3] + src[l4]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
481 sums[4] = src[l4] + src[l5]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
482 sums[5] = src[l5] + src[l6]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
483 sums[6] = src[l6] + src[l7]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
484 sums[7] = src[l7] + src[l8]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
485 sums[8] = src[l8] + last; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
486 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
487 src[l1]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
488 src[l2]= ((src[l2]<<2) + (first + sums[0] + sums[3]<<1) + sums[5] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
489 src[l3]= ((src[l3]<<2) + (first + sums[1] + sums[4]<<1) + sums[6] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
490 src[l4]= ((src[l4]<<2) + (sums[2] + sums[5]<<1) + sums[0] + sums[7] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
491 src[l5]= ((src[l5]<<2) + (sums[3] + sums[6]<<1) + sums[1] + sums[8] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
492 src[l6]= ((src[l6]<<2) + (last + sums[7] + sums[4]<<1) + sums[2] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
493 src[l7]= ((last + src[l7]<<2) + (src[l8] + sums[5]<<1) + sums[3] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
494 src[l8]= ((sums[8]<<2) + (last + sums[6]<<1) + sums[4] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
495 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
496 src++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
497 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
498 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
499 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
500 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
501 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
502 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
|
503 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
504 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
505 src+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
506 //FIXME try pmul for *5 stuff |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
507 // src[0]=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
508 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
509 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
510 "leal (%0, %1), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
511 "leal (%%eax, %1, 4), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
512 // 0 1 2 3 4 5 6 7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
513 // %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
|
514 // %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
|
515 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
516 "movq (%0), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
517 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
518 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
519 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
520 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
521 "movq (%%eax), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
522 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
523 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
524 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
525 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
526 "movq (%%eax, %1), %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
527 "movq %%mm4, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
528 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
529 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
530 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
531 "paddw %%mm0, %%mm0 \n\t" // 2L0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
532 "paddw %%mm1, %%mm1 \n\t" // 2H0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
533 "psubw %%mm4, %%mm2 \n\t" // L1 - L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
534 "psubw %%mm5, %%mm3 \n\t" // H1 - H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
535 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
536 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
537 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
538 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
539 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
540 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
541 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
542 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
543 "movq (%%eax, %1, 2), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
544 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
545 "punpcklbw %%mm7, %%mm2 \n\t" // L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
546 "punpckhbw %%mm7, %%mm3 \n\t" // H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
547 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
548 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
549 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
550 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
551 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
552 "movq %%mm0, temp0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
553 "movq %%mm1, temp1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
554 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
555 "movq (%0, %1, 4), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
556 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
557 "punpcklbw %%mm7, %%mm0 \n\t" // L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
558 "punpckhbw %%mm7, %%mm1 \n\t" // H4 |
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 "psubw %%mm0, %%mm2 \n\t" // L3 - L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
561 "psubw %%mm1, %%mm3 \n\t" // H3 - H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
562 "movq %%mm2, temp2 \n\t" // L3 - L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
563 "movq %%mm3, temp3 \n\t" // H3 - H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
564 "paddw %%mm4, %%mm4 \n\t" // 2L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
565 "paddw %%mm5, %%mm5 \n\t" // 2H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
566 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
567 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
568 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
569 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
570 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
571 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
572 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
573 //50 opcodes so far |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
574 "movq (%%ebx), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
575 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
576 "punpcklbw %%mm7, %%mm2 \n\t" // L5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
577 "punpckhbw %%mm7, %%mm3 \n\t" // H5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
578 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
579 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
580 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
581 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
582 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
583 "movq (%%ebx, %1), %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
584 "punpcklbw %%mm7, %%mm6 \n\t" // L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
585 "psubw %%mm6, %%mm2 \n\t" // L5 - L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
586 "movq (%%ebx, %1), %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
587 "punpckhbw %%mm7, %%mm6 \n\t" // H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
588 "psubw %%mm6, %%mm3 \n\t" // H5 - H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
589 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
590 "paddw %%mm0, %%mm0 \n\t" // 2L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
591 "paddw %%mm1, %%mm1 \n\t" // 2H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
592 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
593 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
594 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
595 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
596 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
597 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
598 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
599 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
600 "movq (%%ebx, %1, 2), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
601 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
602 "punpcklbw %%mm7, %%mm2 \n\t" // L7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
603 "punpckhbw %%mm7, %%mm3 \n\t" // H7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
604 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
605 "paddw %%mm2, %%mm2 \n\t" // 2L7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
606 "paddw %%mm3, %%mm3 \n\t" // 2H7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
607 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
608 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
609 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
610 "movq temp0, %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
611 "movq temp1, %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
612 //FIXME pxor, psubw, pmax for abs |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
613 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
614 "pcmpgtw %%mm0, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
615 "pxor %%mm6, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
616 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
617 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
618 "pcmpgtw %%mm1, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
619 "pxor %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
620 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
621 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
622 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
623 "pcmpgtw %%mm2, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
624 "pxor %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
625 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
626 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
627 "pcmpgtw %%mm3, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
628 "pxor %%mm6, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
629 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
630 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
631 #ifdef HAVE_MMX2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
632 "pminsw %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
633 "pminsw %%mm3, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
634 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
635 "movq %%mm0, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
636 "psubusw %%mm2, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
637 "psubw %%mm6, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
638 "movq %%mm1, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
639 "psubusw %%mm3, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
640 "psubw %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
641 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
642 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
643 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
644 "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
645 "pxor %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
646 "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
647 "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
648 "pxor %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
649 "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
650 // 100 opcodes |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
651 "movd %2, %%mm2 \n\t" // QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
652 //"pcmpeqb %%mm2, %%mm2\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
653 "punpcklwd %%mm2, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
654 "punpcklwd %%mm2, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
655 "psllw $3, %%mm2 \n\t" // 8QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
656 "movq %%mm2, %%mm3 \n\t" // 8QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
657 "pcmpgtw %%mm4, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
658 "pcmpgtw %%mm5, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
659 "pand %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
660 "pand %%mm3, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
661 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
662 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
663 "psubusw %%mm0, %%mm4 \n\t" // hd |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
664 "psubusw %%mm1, %%mm5 \n\t" // ld |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
665 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
666 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
667 "movq w05, %%mm2 \n\t" // 5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
668 "pmullw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
669 "pmullw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
670 "movq w20, %%mm2 \n\t" // 32 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
671 "paddw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
672 "paddw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
673 "psrlw $6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
674 "psrlw $6, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
675 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
676 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
677 "movq w06, %%mm2 \n\t" // 6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
678 "paddw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
679 "paddw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
680 "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
|
681 //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
|
682 "pmulhw %%mm2, %%mm4 \n\t" // hd/13 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
683 "pmulhw %%mm2, %%mm5 \n\t" // ld/13 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
684 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
685 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
686 "movq temp2, %%mm0 \n\t" // L3 - L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
687 "movq temp3, %%mm1 \n\t" // H3 - H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
688 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
689 "pxor %%mm2, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
690 "pxor %%mm3, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
691 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
692 // FIXME rounding error |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
693 "psraw $1, %%mm0 \n\t" // (L3 - L4)/2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
694 "psraw $1, %%mm1 \n\t" // (H3 - H4)/2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
695 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
696 "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
697 "pxor %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
698 "pxor %%mm3, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
699 "psubw %%mm2, %%mm0 \n\t" // |L3-L4| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
700 "psubw %%mm3, %%mm1 \n\t" // |H3-H4| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
701 // "psrlw $1, %%mm0 \n\t" // |L3 - L4|/2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
702 // "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
703 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
704 "pxor %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
705 "pxor %%mm7, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
706 "pand %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
707 "pand %%mm3, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
708 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
709 #ifdef HAVE_MMX2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
710 "pminsw %%mm0, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
711 "pminsw %%mm1, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
712 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
713 "movq %%mm4, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
714 "psubusw %%mm0, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
715 "psubw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
716 "movq %%mm5, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
717 "psubusw %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
718 "psubw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
719 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
720 "pxor %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
721 "pxor %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
722 "psubw %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
723 "psubw %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
724 "packsswb %%mm5, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
725 "movq (%%eax, %1, 2), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
726 "paddb %%mm4, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
727 "movq %%mm0, (%%eax, %1, 2) \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
728 "movq (%0, %1, 4), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
729 "psubb %%mm4, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
730 // "pxor %%mm0, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
731 "movq %%mm0, (%0, %1, 4) \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
732 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
733 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
734 : "r" (src), "r" (stride), "r" (QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
735 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
736 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
737 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
738 const int l1= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
739 const int l2= stride + l1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
740 const int l3= stride + l2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
741 const int l4= stride + l3; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
742 const int l5= stride + l4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
743 const int l6= stride + l5; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
744 const int l7= stride + l6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
745 const int l8= stride + l7; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
746 // const int l9= stride + l8; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
747 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
748 for(int x=0; x<BLOCK_SIZE; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
749 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
750 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
|
751 if(ABS(middleEnergy) < 8*QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
752 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
753 const int q=(src[l4] - src[l5])/2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
754 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
|
755 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
|
756 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
757 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
758 d= MAX(d, 0); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
759 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
760 d= (5*d + 32) >> 6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
761 d*= SIGN(-middleEnergy); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
762 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
763 if(q>0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
764 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
765 d= d<0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
766 d= d>q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
767 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
768 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
769 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
770 d= d>0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
771 d= d<q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
772 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
773 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
774 src[l4]-= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
775 src[l5]+= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
776 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
777 src++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
778 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
779 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
780 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
781 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
782 //FIXME? |255-0| = 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
783 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
784 * Check if the given 8x8 Block is mostly "flat" and copy the unaliged data into tempBlock. |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
785 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
786 static inline bool isHorizDCAndCopy2Temp(uint8_t src[], int stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
787 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
788 // src++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
789 int numEq= 0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
790 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
791 asm volatile ( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
792 // "int $3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
793 "pushl %1\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
794 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
795 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
796 "leal tempBlock, %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
797 "pxor %%mm0, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
798 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
799 #define HDC_CHECK_AND_CPY(i) \ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
800 "movq -4(%1), %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
801 "psrlq $32, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
802 "punpckldq 4(%1), %%mm2 \n\t" /* (%1) */\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
803 "movq %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
804 "psrlq $8, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
805 "psubb %%mm1, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
806 "paddb %%mm7, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
807 "pcmpgtb %%mm6, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
808 "paddb %%mm2, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
809 "movq %%mm1," #i "(%%eax) \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
810 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
811 HDC_CHECK_AND_CPY(0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
812 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
813 HDC_CHECK_AND_CPY(8) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
814 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
815 HDC_CHECK_AND_CPY(16) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
816 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
817 HDC_CHECK_AND_CPY(24) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
818 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
819 HDC_CHECK_AND_CPY(32) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
820 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
821 HDC_CHECK_AND_CPY(40) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
822 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
823 HDC_CHECK_AND_CPY(48) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
824 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
825 HDC_CHECK_AND_CPY(56) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
826 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
827 "psllq $8, %%mm0 \n\t" // remove dummy value |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
828 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
829 "psrlw $8, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
830 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
831 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
832 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
833 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
834 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
835 "psrlq $32, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
836 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
837 "popl %1\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
838 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
839 : "=r" (numEq) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
840 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
841 : "%eax" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
842 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
843 // printf("%d\n", numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
844 numEq= (256 - (numEq & 0xFF)) &0xFF; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
845 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
846 for(int y=0; y<BLOCK_SIZE; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
847 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
848 if(((src[0] - src[1] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
849 if(((src[1] - src[2] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
850 if(((src[2] - src[3] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
851 if(((src[3] - src[4] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
852 if(((src[4] - src[5] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
853 if(((src[5] - src[6] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
854 if(((src[6] - src[7] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
855 tempBlock[0 + y*TEMP_STRIDE] = src[0]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
856 tempBlock[1 + y*TEMP_STRIDE] = src[1]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
857 tempBlock[2 + y*TEMP_STRIDE] = src[2]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
858 tempBlock[3 + y*TEMP_STRIDE] = src[3]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
859 tempBlock[4 + y*TEMP_STRIDE] = src[4]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
860 tempBlock[5 + y*TEMP_STRIDE] = src[5]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
861 tempBlock[6 + y*TEMP_STRIDE] = src[6]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
862 tempBlock[7 + y*TEMP_STRIDE] = src[7]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
863 src+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
864 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
865 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
866 /* if(abs(numEq - asmEq) > 0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
867 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
868 // printf("\nasm:%d c:%d\n", asmEq, numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
869 for(int y=0; y<8; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
870 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
871 for(int x=0; x<8; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
872 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
873 printf("%d ", src[x + y*stride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
874 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
875 printf("\n"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
876 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
877 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
878 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
879 // printf("%d\n", numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
880 return numEq > hFlatnessThreshold; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
881 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
882 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
883 static inline bool isHorizMinMaxOk(uint8_t src[], int stride, int QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
884 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
885 #ifdef MMX_FIXME |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
886 FIXME |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
887 int isOk; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
888 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
889 // "int $3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
890 "movq (%1, %2), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
891 "movq (%1, %2, 8), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
892 "movq %%mm0, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
893 "psubusb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
894 "psubusb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
895 "por %%mm1, %%mm0 \n\t" // ABS Diff |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
896 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
897 "movq pQPb, %%mm7 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
898 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
899 "psubusb %%mm7, %%mm0 \n\t" // Diff <= 2QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
900 "pcmpeqd b00, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
901 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
902 "pcmpeqd bFF, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
903 // "movd %%mm0, (%1, %2, 4)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
904 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
905 : "=r" (isOk) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
906 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
907 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
908 return isOk; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
909 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
910 if(abs(src[0] - src[7]) > 2*QP) return false; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
911 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
912 return true; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
913 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
914 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
915 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
916 static inline void doHorizDefFilterAndCopyBack(uint8_t dst[], int stride, int QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
917 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
918 #ifdef HAVE_MMX2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
919 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
920 "pushl %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
921 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
922 "movq bm00001000, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
923 "movd %2, %%mm5 \n\t" // QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
924 "movq %%mm5, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
925 "paddusb %%mm5, %%mm5 \n\t" // 2QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
926 "paddusb %%mm5, %%mm4 \n\t" // 3QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
927 "psllq $24, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
928 "pxor %%mm5, %%mm5 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
929 "psubb %%mm4, %%mm5 \n\t" // -QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
930 "leal tempBlock, %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
931 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
932 //FIXME? "unroll by 2" and mix |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
933 #define HDF(i) "movq " #i "(%%eax), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
934 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
935 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
936 "psrlq $8, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
937 "psubusb %%mm1, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
938 "psubusb %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
939 "por %%mm2, %%mm1 \n\t" /* |px - p(x+1)| */\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
940 "pcmpeqb %%mm7, %%mm2 \n\t" /* sgn[px - p(x+1)] */\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
941 "pshufw $0xAA, %%mm1, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
942 "pminub %%mm1, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
943 "psrlq $16, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
944 "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
|
945 "paddb %%mm5, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
946 "psubusb %%mm5, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
947 "psrlw $2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
948 "pxor %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
949 "psubb %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
950 "pand %%mm6, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
951 "psubb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
952 "psllq $8, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
953 "paddb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
954 "movd %%mm0, (%0) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
955 "psrlq $32, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
956 "movd %%mm0, 4(%0) \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
957 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
958 HDF(0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
959 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
960 HDF(8) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
961 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
962 HDF(16) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
963 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
964 HDF(24) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
965 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
966 HDF(32) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
967 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
968 HDF(40) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
969 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
970 HDF(48) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
971 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
972 HDF(56) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
973 "popl %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
974 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
975 : "r" (dst), "r" (stride), "r" (QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
976 : "%eax" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
977 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
978 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
979 uint8_t *src= tempBlock; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
980 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
981 for(int y=0; y<BLOCK_SIZE; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
982 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
983 dst[0] = src[0]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
984 dst[1] = src[1]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
985 dst[2] = src[2]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
986 dst[3] = src[3]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
987 dst[4] = src[4]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
988 dst[5] = src[5]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
989 dst[6] = src[6]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
990 dst[7] = src[7]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
991 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
992 const int middleEnergy= 5*(src[4] - src[5]) + 2*(src[2] - src[5]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
993 if(ABS(middleEnergy) < 8*QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
994 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
995 const int q=(src[3] - src[4])/2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
996 const int leftEnergy= 5*(src[2] - src[1]) + 2*(src[0] - src[3]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
997 const int rightEnergy= 5*(src[6] - src[5]) + 2*(src[4] - src[7]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
998 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
999 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1000 d= MAX(d, 0); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1001 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1002 d= (5*d + 32) >> 6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1003 d*= SIGN(-middleEnergy); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1004 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1005 if(q>0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1006 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1007 d= d<0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1008 d= d>q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1009 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1010 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1011 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1012 d= d>0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1013 d= d<q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1014 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1015 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1016 dst[3]-= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1017 dst[4]+= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1018 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1019 dst+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1020 src+= TEMP_STRIDE; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1021 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1022 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1023 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1024 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1025 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1026 * Do a horizontal low pass filter on the 8x8 block |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1027 * useing the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1028 * useing approximately the 7-Tap Filter (1,2,3,4,3,2,1)/16 (MMX2 version) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1029 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1030 static inline void doHorizLowPassAndCopyBack(uint8_t dst[], int stride, int QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1031 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1032 //return; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1033 #ifdef HAVE_MMX2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1034 asm volatile( //"movv %0 %1 %2\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1035 "pushl %0\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1036 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1037 "leal tempBlock, %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1038 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1039 #define HLP1 "movq (%0), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1040 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1041 "psllq $8, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1042 "pavgb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1043 "psrlw $8, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1044 "pxor %%mm1, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1045 "packuswb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1046 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1047 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1048 "psllq $32, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1049 "paddb %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1050 "psllq $16, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1051 "pavgb %%mm2, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1052 "movq %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1053 "pand bm11001100, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1054 "paddusb %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1055 "psrlq $8, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1056 "pavgb %%mm1, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1057 "pavgb %%mm3, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1058 "psrlq $16, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1059 "punpcklbw %%mm2, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1060 "movq %%mm2, (%0) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1061 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1062 #define HLP2 "movq (%0), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1063 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1064 "psllq $8, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1065 "pavgb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1066 "psrlw $8, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1067 "pxor %%mm1, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1068 "packuswb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1069 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1070 "psllq $32, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1071 "psllq $16, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1072 "pavgb %%mm2, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1073 "movq %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1074 "pand bm11001100, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1075 "paddusb %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1076 "psrlq $8, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1077 "pavgb %%mm3, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1078 "psrlq $16, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1079 "punpcklbw %%mm2, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1080 "movq %%mm2, (%0) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1081 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1082 // 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
|
1083 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1084 31 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1085 121 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1086 121 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1087 121 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1088 121 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1089 121 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1090 121 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1091 13 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1092 Implemented Exact 7-Tap |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1093 9421 A321 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1094 36421 64321 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1095 334321 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1096 1234321 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1097 1234321 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1098 123433 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1099 12463 12346 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1100 1249 123A |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1101 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1102 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1103 #define HLP3(i) "movq " #i "(%%eax), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1104 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1105 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1106 "movq %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1107 "movq %%mm0, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1108 "psllq $8, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1109 "psrlq $8, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1110 "pand bm00000001, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1111 "pand bm10000000, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1112 "por %%mm3, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1113 "por %%mm4, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1114 "pavgb %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1115 "pavgb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1116 \ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1117 "pshufw $0xF9, %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1118 "pshufw $0x90, %%mm0, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1119 "pavgb %%mm3, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1120 "pavgb %%mm4, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1121 "movd %%mm0, (%0) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1122 "psrlq $32, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1123 "movd %%mm0, 4(%0) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1124 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1125 #define HLP(i) HLP3(i) |
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 HLP(0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1128 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1129 HLP(8) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1130 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1131 HLP(16) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1132 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1133 HLP(24) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1134 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1135 HLP(32) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1136 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1137 HLP(40) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1138 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1139 HLP(48) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1140 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1141 HLP(56) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1142 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1143 "popl %0\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1144 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1145 : "r" (dst), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1146 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1147 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1148 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1149 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1150 uint8_t *temp= tempBlock; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1151 for(int y=0; y<BLOCK_SIZE; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1152 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1153 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
|
1154 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
|
1155 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1156 int sums[9]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1157 sums[0] = first + temp[0]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1158 sums[1] = temp[0] + temp[1]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1159 sums[2] = temp[1] + temp[2]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1160 sums[3] = temp[2] + temp[3]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1161 sums[4] = temp[3] + temp[4]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1162 sums[5] = temp[4] + temp[5]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1163 sums[6] = temp[5] + temp[6]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1164 sums[7] = temp[6] + temp[7]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1165 sums[8] = temp[7] + last; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1166 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1167 dst[0]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1168 dst[1]= ((dst[1]<<2) + (first + sums[0] + sums[3]<<1) + sums[5] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1169 dst[2]= ((dst[2]<<2) + (first + sums[1] + sums[4]<<1) + sums[6] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1170 dst[3]= ((dst[3]<<2) + (sums[2] + sums[5]<<1) + sums[0] + sums[7] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1171 dst[4]= ((dst[4]<<2) + (sums[3] + sums[6]<<1) + sums[1] + sums[8] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1172 dst[5]= ((dst[5]<<2) + (last + sums[7] + sums[4]<<1) + sums[2] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1173 dst[6]= ((last + dst[6]<<2) + (dst[7] + sums[5]<<1) + sums[3] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1174 dst[7]= ((sums[8]<<2) + (last + sums[6]<<1) + sums[4] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1175 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1176 dst+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1177 temp+= TEMP_STRIDE; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1178 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1179 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1180 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1181 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1182 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1183 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
|
1184 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1185 //FIXME |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1186 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1187 #ifdef HAVE_MMX2X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1188 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1189 "leal (%0, %1), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1190 "leal (%%eax, %1, 4), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1191 // 0 1 2 3 4 5 6 7 8 9 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1192 // %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
|
1193 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1194 "pcmpeq %%mm6, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1195 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1196 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1197 #define FIND_MIN_MAX(addr)\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1198 "movq (" #addr "), %%mm0, \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1199 "pminub %%mm0, %%mm6 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1200 "pmaxub %%mm0, %%mm7 \n\t" |
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 FIND_MIN_MAX(%0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1203 FIND_MIN_MAX(%%eax) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1204 FIND_MIN_MAX(%%eax, %1) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1205 FIND_MIN_MAX(%%eax, %1, 2) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1206 FIND_MIN_MAX(%0, %1, 4) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1207 FIND_MIN_MAX(%%ebx) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1208 FIND_MIN_MAX(%%ebx, %1) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1209 FIND_MIN_MAX(%%ebx, %1, 2) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1210 FIND_MIN_MAX(%0, %1, 8) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1211 FIND_MIN_MAX(%%ebx, %1, 2) |
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 %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1214 "psrlq $32, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1215 "pminub %%mm4, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1216 "movq %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1217 "psrlq $16, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1218 "pminub %%mm4, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1219 "movq %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1220 "psrlq $8, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1221 "pminub %%mm4, %%mm6 \n\t" // min of pixels |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1222 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1223 "movq %%mm7, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1224 "psrlq $32, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1225 "pmaxub %%mm4, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1226 "movq %%mm7, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1227 "psrlq $16, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1228 "pmaxub %%mm4, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1229 "movq %%mm7, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1230 "psrlq $8, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1231 "pmaxub %%mm4, %%mm7 \n\t" // max of pixels |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1232 "pavgb %%mm6, %%mm7 \n\t" // (max + min)/2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1233 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1234 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1235 : : "r" (src), "r" (stride), "r" (QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1236 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1237 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1238 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1239 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1240 //FIXME |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1241 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1242 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1243 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1244 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1245 * ... |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1246 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1247 extern "C"{ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1248 void postprocess(unsigned char * src[], int src_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1249 unsigned char * dst[], int dst_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1250 int horizontal_size, int vertical_size, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1251 QP_STORE_T *QP_store, int QP_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1252 int mode) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1253 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1254 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1255 long long T= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1256 for(int y=vertical_size-1; y>=0 ; y--) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1257 memcpy(dst[0] + y*src_stride, src[0] + y*src_stride,src_stride); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1258 // memcpy(dst[0], src[0],src_stride*vertical_size); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1259 printf("%4dk\r", (rdtsc()-T)/1000); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1260 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1261 return; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1262 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1263 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1264 long long T= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1265 while( (rdtsc() - T)/1000 < 4000); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1266 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1267 return; |
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 postProcess(src[0], src_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1270 dst[0], dst_stride, horizontal_size, vertical_size, QP_store, QP_stride, false); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1271 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1272 horizontal_size >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1273 vertical_size >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1274 src_stride >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1275 dst_stride >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1276 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1277 if(1) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1278 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1279 postProcess(src[1], src_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1280 dst[1], dst_stride, horizontal_size, vertical_size, QP_store, QP_stride, true); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1281 postProcess(src[2], src_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1282 dst[2], dst_stride, horizontal_size, vertical_size, QP_store, QP_stride, true); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1283 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1284 else |
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 memcpy(dst[1], src[1], src_stride*horizontal_size); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1287 memcpy(dst[2], src[2], src_stride*horizontal_size); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1288 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1289 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1290 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1291 |
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 * Copies a block from src to dst and fixes the blacklevel |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1294 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1295 static inline void blockCopy(uint8_t dst[], int dstStride, uint8_t src[], int srcStride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1296 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1297 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1298 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1299 "pushl %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1300 "pushl %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1301 "leal (%2,%2), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1302 "leal (%3,%3), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1303 "movq packedYOffset, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1304 "movq packedYScale, %%mm3 \n\t" |
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 #define SIMPLE_CPY \ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1307 "movq (%0), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1308 "movq (%0,%2), %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1309 "psubusb %%mm2, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1310 "psubusb %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1311 "movq %%mm0, (%1) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1312 "movq %%mm1, (%1, %3) \n\t"\ |
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 #define SCALED_CPY \ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1315 "movq (%0), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1316 "movq (%0,%2), %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1317 "psubusb %%mm2, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1318 "psubusb %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1319 "pxor %%mm4, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1320 "pxor %%mm5, %%mm5 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1321 "punpcklbw %%mm0, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1322 "punpckhbw %%mm0, %%mm5 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1323 "pmulhuw %%mm3, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1324 "pmulhuw %%mm3, %%mm5 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1325 "packuswb %%mm5, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1326 "movq %%mm4, (%1) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1327 "pxor %%mm4, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1328 "pxor %%mm5, %%mm5 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1329 "punpcklbw %%mm1, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1330 "punpckhbw %%mm1, %%mm5 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1331 "pmulhuw %%mm3, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1332 "pmulhuw %%mm3, %%mm5 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1333 "packuswb %%mm5, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1334 "movq %%mm4, (%1, %3) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1335 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1336 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1337 #define CPY SCALED_CPY |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1338 //#define CPY SIMPLE_CPY |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1339 // "prefetchnta 8(%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1340 CPY |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1341 "addl %%eax, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1342 "addl %%ebx, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1343 CPY |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1344 "addl %%eax, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1345 "addl %%ebx, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1346 CPY |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1347 "addl %%eax, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1348 "addl %%ebx, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1349 CPY |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1350 "popl %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1351 "popl %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1352 : : "r" (src), |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1353 "r" (dst), |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1354 "r" (srcStride), |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1355 "r" (dstStride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1356 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1357 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1358 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1359 for(int i=0; i<BLOCK_SIZE; i++) // last 10x8 Block is copied allready so +2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1360 memcpy( &(dst[dstStride*i]), |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1361 &(src[srcStride*i]), BLOCK_SIZE); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1362 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1363 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1364 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1365 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1366 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1367 * Filters array of bytes (Y or U or V values) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1368 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1369 void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1370 QP_STORE_T QPs[], int QPStride, bool isColor) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1371 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1372 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1373 #ifdef TIMEING |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1374 long long T0, T1, memcpyTime=0, vertTime=0, horizTime=0, sumTime, diffTime=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1375 sumTime= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1376 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1377 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1378 /* we need 64bit here otherwise weŽll going to have a problem |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1379 after watching a black picture for 5 hours*/ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1380 static uint64_t *yHistogram= NULL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1381 if(!yHistogram) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1382 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1383 yHistogram= new uint64_t[256]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1384 for(int i=0; i<256; i++) yHistogram[i]= width*height/64/256; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1385 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1386 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1387 int black=0, white=255; // blackest black and whitest white in the picture |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1388 if(!isColor) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1389 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1390 uint64_t sum= 0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1391 for(int i=0; i<256; i++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1392 sum+= yHistogram[i]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1393 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1394 uint64_t maxClipped= (uint64_t)(sum * maxClippedThreshold); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1395 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1396 uint64_t clipped= sum; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1397 for(black=255; black>0; black--) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1398 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1399 if(clipped < maxClipped) break; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1400 clipped-= yHistogram[black]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1401 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1402 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1403 clipped= sum; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1404 for(white=0; white<256; white++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1405 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1406 if(clipped < maxClipped) break; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1407 clipped-= yHistogram[white]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1408 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1409 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1410 // we cant handle negative correctures |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1411 packedYOffset= MAX(black - minAllowedY, 0); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1412 packedYOffset|= packedYOffset<<32; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1413 packedYOffset|= packedYOffset<<16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1414 packedYOffset|= packedYOffset<<8; |
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 // uint64_t scale= (int)(256.0*256.0/(white-black) + 0.5); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1417 double scale= (double)(maxAllowedY - minAllowedY) / (double)(white-black); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1418 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1419 packedYScale= uint16_t(scale*256.0 + 0.5); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1420 packedYScale|= packedYScale<<32; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1421 packedYScale|= packedYScale<<16; |
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 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1424 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1425 packedYScale= 0x0100010001000100LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1426 packedYOffset= 0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1427 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1428 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1429 for(int x=0; x<width; x+=BLOCK_SIZE) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1430 blockCopy(dst + x, dstStride, src + x, srcStride); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1431 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1432 for(int y=0; y<height; y+=BLOCK_SIZE) |
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 //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
|
1435 uint8_t *srcBlock= &(src[y*srcStride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1436 uint8_t *dstBlock= &(dst[y*dstStride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1437 uint8_t *vertSrcBlock= &(srcBlock[srcStride*3]); // Blocks are 10x8 -> *3 to start |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1438 uint8_t *vertBlock= &(dstBlock[dstStride*3]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1439 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1440 // finish 1 block before the next otherwise weŽll might have a problem |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1441 // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1442 for(int x=0; x<width; x+=BLOCK_SIZE) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1443 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1444 int QP= isColor ? |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1445 QPs[(y>>3)*QPStride + (x>>3)]: |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1446 (QPs[(y>>4)*QPStride + (x>>4)] * (packedYScale &0xFFFF))>>8; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1447 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1448 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1449 "movd %0, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1450 "packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1451 "packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1452 "packuswb %%mm7, %%mm7 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1453 "movq %%mm7, pQPb \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1454 : : "r" (QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1455 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1456 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1457 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1458 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1459 const int stride= dstStride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1460 if(y + 12 < height) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1461 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1462 #ifdef MORE_TIMEING |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1463 T0= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1464 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1465 #ifdef HAVE_MMX2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1466 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1467 prefetchnta(vertSrcBlock + (((x>>3)&3) + 2)*srcStride + 32); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1468 prefetchnta(vertSrcBlock + (((x>>3)&3) + 6)*srcStride + 32); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1469 prefetcht0(vertBlock + (((x>>3)&3) + 2)*dstStride + 32); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1470 prefetcht0(vertBlock + (((x>>3)&3) + 6)*dstStride + 32); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1471 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1472 if(!isColor) yHistogram[ srcBlock[0] ]++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1473 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1474 blockCopy(vertBlock + dstStride*2, dstStride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1475 vertSrcBlock + srcStride*2, srcStride); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1476 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1477 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1478 #ifdef MORE_TIMEING |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1479 T1= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1480 memcpyTime+= T1-T0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1481 T0=T1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1482 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1483 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1484 if( isVertDC(vertBlock, stride)) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1485 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1486 if(isVertMinMaxOk(vertBlock, stride, QP)) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1487 doVertLowPass(vertBlock, stride, QP); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1488 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1489 else if(x<width) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1490 doVertDefFilter(vertBlock, stride, QP); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1491 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1492 #ifdef MORE_TIMEING |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1493 T1= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1494 vertTime+= T1-T0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1495 T0=T1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1496 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1497 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1498 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1499 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1500 for(int i=2; i<BLOCK_SIZE/2+1; i++) // last 10x8 Block is copied allready so +2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1501 memcpy( &(vertBlock[dstStride*i]), |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1502 &(vertSrcBlock[srcStride*i]), BLOCK_SIZE); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1503 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1504 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1505 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1506 if(x - 8 >= 0 && x<width) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1507 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1508 #ifdef MORE_TIMEING |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1509 T0= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1510 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1511 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1512 if( isHorizDCAndCopy2Temp(dstBlock-4, stride)) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1513 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1514 if(isHorizMinMaxOk(tempBlock, TEMP_STRIDE, QP)) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1515 doHorizLowPassAndCopyBack(dstBlock-4, stride, QP); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1516 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1517 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1518 doHorizDefFilterAndCopyBack(dstBlock-4, stride, QP); |
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 #ifdef MORE_TIMEING |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1521 T1= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1522 horizTime+= T1-T0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1523 T0=T1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1524 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1525 dering(dstBlock - 9 - stride, stride, QP); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1526 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1527 else if(y!=0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1528 dering(dstBlock - stride*9 + width-9, stride, QP); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1529 //FIXME dering filter will not be applied to last block (bottom right) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1530 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1531 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1532 dstBlock+=8; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1533 srcBlock+=8; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1534 vertBlock+=8; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1535 vertSrcBlock+=8; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1536 } |
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 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1539 asm volatile("emms"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1540 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1541 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1542 #ifdef TIMEING |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1543 // 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
|
1544 sumTime= rdtsc() - sumTime; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1545 if(!isColor) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1546 printf("cpy:%4dk, vert:%4dk, horiz:%4dk, sum:%4dk, diff:%4dk, color: %d/%d \r", |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1547 int(memcpyTime/1000), int(vertTime/1000), int(horizTime/1000), |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1548 int(sumTime/1000), int((sumTime-memcpyTime-vertTime-horizTime)/1000) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1549 , black, white); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1550 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1551 } |