Mercurial > mplayer.hg
annotate libmpcodecs/vf_fspp.c @ 15633:524d306ce0f3
disable mmx code for x86-64
author | henry |
---|---|
date | Sat, 04 Jun 2005 12:41:19 +0000 |
parents | e813a3e431a8 |
children | 7eddcf69a5fd |
rev | line source |
---|---|
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1 /* |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2 Copyright (C) 2003 Michael Niedermayer <michaelni@gmx.at> |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
3 Copyright (C) 2005 Nikolaj Poroshin <porosh3@psu.ru> |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
5 This program is free software; you can redistribute it and/or modify |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
6 it under the terms of the GNU General Public License as published by |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
7 the Free Software Foundation; either version 2 of the License, or |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
8 (at your option) any later version. |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
9 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
10 This program is distributed in the hope that it will be useful, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
11 but WITHOUT ANY WARRANTY; without even the implied warranty of |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
13 GNU General Public License for more details. |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
14 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
15 You should have received a copy of the GNU General Public License |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
16 along with this program; if not, write to the Free Software |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
17 Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
18 */ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
19 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
20 /* |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
21 * This implementation is based on an algorithm described in |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
22 * "Aria Nosratinia Embedded Post-Processing for |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
23 * Enhancement of Compressed Images (1999)" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
24 * (http://citeseer.nj.nec.com/nosratinia99embedded.html) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
25 * Futher, with splitting (i)dct into hor/ver passes, one of them can be |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
26 * performed once per block, not pixel. This allows for much better speed. |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
27 */ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
28 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
29 /* |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
30 Heavily optimized version of SPP filter by Nikolaj |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
31 */ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
32 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
33 #include <stdio.h> |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
34 #include <stdlib.h> |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
35 #include <string.h> |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
36 #include <inttypes.h> |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
37 #include <math.h> |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
38 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
39 #include "../config.h" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
40 |
15633 | 41 #ifdef ARCH_X86_64 |
42 // until the mmx code is fixed to support x86-64 | |
43 #undef HAVE_MMX | |
44 #endif | |
45 | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
46 #ifdef USE_LIBAVCODEC |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
47 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
48 #include "../mp_msg.h" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
49 #include "../cpudetect.h" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
50 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
51 #ifdef USE_LIBAVCODEC_SO |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
52 #include <ffmpeg/avcodec.h> |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
53 #include <ffmpeg/dsputil.h> |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
54 #else |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
55 #include "../libavcodec/avcodec.h" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
56 #include "../libavcodec/dsputil.h" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
57 #endif |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
58 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
59 #ifdef HAVE_MALLOC_H |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
60 #include <malloc.h> |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
61 #endif |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
62 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
63 #include "img_format.h" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
64 #include "mp_image.h" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
65 #include "vf.h" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
66 #include "../libvo/fastmemcpy.h" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
67 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
68 //===========================================================================// |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
69 #define BLOCKSZ 12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
70 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
71 static const short custom_threshold[64]= |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
72 // values (296) can't be too high |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
73 // -it causes too big quant dependence |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
74 // or maybe overflow(check), which results in some flashing |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
75 { 71, 296, 295, 237, 71, 40, 38, 19, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
76 245, 193, 185, 121, 102, 73, 53, 27, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
77 158, 129, 141, 107, 97, 73, 50, 26, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
78 102, 116, 109, 98, 82, 66, 45, 23, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
79 71, 94, 95, 81, 70, 56, 38, 20, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
80 56, 77, 74, 66, 56, 44, 30, 15, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
81 38, 53, 50, 45, 38, 30, 21, 11, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
82 20, 27, 26, 23, 20, 15, 11, 5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
83 }; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
84 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
85 static const uint8_t __attribute__((aligned(32))) dither[8][8]={ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
86 { 0, 48, 12, 60, 3, 51, 15, 63, }, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
87 { 32, 16, 44, 28, 35, 19, 47, 31, }, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
88 { 8, 56, 4, 52, 11, 59, 7, 55, }, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
89 { 40, 24, 36, 20, 43, 27, 39, 23, }, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
90 { 2, 50, 14, 62, 1, 49, 13, 61, }, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
91 { 34, 18, 46, 30, 33, 17, 45, 29, }, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
92 { 10, 58, 6, 54, 9, 57, 5, 53, }, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
93 { 42, 26, 38, 22, 41, 25, 37, 21, }, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
94 }; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
95 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
96 struct vf_priv_s { //align 16 ! |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
97 uint64_t threshold_mtx_noq[8*2]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
98 uint64_t threshold_mtx[8*2];//used in both C & MMX (& later SSE2) versions |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
99 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
100 int log2_count; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
101 int temp_stride; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
102 int qp; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
103 int mpeg2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
104 int prev_q; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
105 uint8_t *src; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
106 int16_t *temp; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
107 //int mode; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
108 }; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
109 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
110 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
111 #ifndef HAVE_MMX |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
112 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
113 //This func reads from 1 slice, 1 and clears 0 & 1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
114 static void store_slice_c(uint8_t *dst, int16_t *src, int dst_stride, int src_stride, int width, int height, int log2_scale) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
115 {int y, x; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
116 #define STORE(pos) \ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
117 temp= (src[x + pos] + (d[pos]>>log2_scale))>>(6-log2_scale); \ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
118 src[x + pos]=src[x + pos - 8*src_stride]=0; \ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
119 if(temp & 0x100) temp= ~(temp>>31); \ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
120 dst[x + pos]= temp; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
121 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
122 for(y=0; y<height; y++){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
123 const uint8_t *d= dither[y]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
124 for(x=0; x<width; x+=8){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
125 int temp; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
126 STORE(0); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
127 STORE(1); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
128 STORE(2); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
129 STORE(3); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
130 STORE(4); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
131 STORE(5); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
132 STORE(6); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
133 STORE(7); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
134 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
135 src+=src_stride; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
136 dst+=dst_stride; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
137 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
138 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
139 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
140 //This func reads from 2 slices, 0 & 2 and clears 2-nd |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
141 static void store_slice2_c(uint8_t *dst, int16_t *src, int dst_stride, int src_stride, int width, int height, int log2_scale) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
142 {int y, x; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
143 #define STORE2(pos) \ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
144 temp= (src[x + pos] + src[x + pos + 16*src_stride] + (d[pos]>>log2_scale))>>(6-log2_scale); \ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
145 src[x + pos + 16*src_stride]=0; \ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
146 if(temp & 0x100) temp= ~(temp>>31); \ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
147 dst[x + pos]= temp; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
148 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
149 for(y=0; y<height; y++){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
150 const uint8_t *d= dither[y]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
151 for(x=0; x<width; x+=8){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
152 int temp; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
153 STORE2(0); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
154 STORE2(1); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
155 STORE2(2); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
156 STORE2(3); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
157 STORE2(4); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
158 STORE2(5); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
159 STORE2(6); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
160 STORE2(7); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
161 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
162 src+=src_stride; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
163 dst+=dst_stride; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
164 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
165 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
166 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
167 static void mul_thrmat_c(struct vf_priv_s *p,int q) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
168 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
169 int a; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
170 for(a=0;a<64;a++) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
171 ((short*)p->threshold_mtx)[a]=q * ((short*)p->threshold_mtx_noq)[a];//ints faster in C |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
172 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
173 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
174 static void column_fidct_c(int16_t* thr_adr, DCTELEM *data, DCTELEM *output, int cnt); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
175 static void row_idct_c(DCTELEM* workspace, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
176 int16_t* output_adr, int output_stride, int cnt); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
177 static void row_fdct_c(DCTELEM *data, const uint8_t *pixels, int line_size, int cnt); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
178 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
179 //this is rather ugly, but there is no need for function pointers |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
180 #define store_slice_s store_slice_c |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
181 #define store_slice2_s store_slice2_c |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
182 #define mul_thrmat_s mul_thrmat_c |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
183 #define column_fidct_s column_fidct_c |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
184 #define row_idct_s row_idct_c |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
185 #define row_fdct_s row_fdct_c |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
186 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
187 #else /* HAVE_MMX */ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
188 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
189 //This func reads from 1 slice, 1 and clears 0 & 1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
190 static void store_slice_mmx(uint8_t *dst, int16_t *src, int dst_stride, int src_stride, int width, int height, int log2_scale) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
191 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
192 const uint8_t *od=&dither[0][0]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
193 width = (width+7)&~7; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
194 dst_stride-=width; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
195 //src_stride=(src_stride-width)*2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
196 height=(int)(&dither[height][0]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
197 asm volatile( |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
198 "movl %5, %%edx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
199 "movl %6, %%esi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
200 "movl %7, %%edi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
201 "movl %1, %%eax \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
202 "movd %%edx, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
203 "xorl $-1, %%edx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
204 "movl %%eax, %%ecx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
205 "addl $7, %%edx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
206 "negl %%eax \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
207 "subl %0, %%ecx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
208 "addl %%ecx, %%ecx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
209 "movd %%edx, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
210 "movl %%ecx, %1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
211 "movl %2, %%edx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
212 "shll $4, %%eax \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
213 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
214 "2: \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
215 "movq (%%edx), %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
216 "movq %%mm3, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
217 "pxor %%mm7, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
218 "punpcklbw %%mm7, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
219 "punpckhbw %%mm7, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
220 "movl %0, %%ecx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
221 "psraw %%mm5, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
222 "psraw %%mm5, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
223 "1: \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
224 "movq %%mm7, (%%esi,%%eax,) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
225 "movq (%%esi), %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
226 "movq 8(%%esi), %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
227 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
228 "movq %%mm7, 8(%%esi,%%eax,) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
229 "paddw %%mm3, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
230 "paddw %%mm4, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
231 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
232 "movq %%mm7, (%%esi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
233 "psraw %%mm2, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
234 "psraw %%mm2, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
235 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
236 "movq %%mm7, 8(%%esi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
237 "packuswb %%mm1, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
238 "addl $16, %%esi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
239 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
240 "movq %%mm0, (%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
241 "addl $8, %%edi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
242 "subl $8, %%ecx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
243 "jg 1b \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
244 "addl %1, %%esi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
245 "addl $8, %%edx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
246 "addl %3, %%edi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
247 "cmpl %4, %%edx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
248 "jl 2b \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
249 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
250 : |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
251 : "m" (width), "m" (src_stride), "g" (od), "m" (dst_stride), "g" (height), |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
252 "m" (log2_scale), "m" (src), "m" (dst) //input |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
253 : "%eax", "%ecx", "%edx", "%esi", "%edi" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
254 ); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
255 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
256 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
257 //This func reads from 2 slices, 0 & 2 and clears 2-nd |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
258 static void store_slice2_mmx(uint8_t *dst, int16_t *src, int dst_stride, int src_stride, int width, int height, int log2_scale) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
259 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
260 const uint8_t *od=&dither[0][0]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
261 width = (width+7)&~7; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
262 dst_stride-=width; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
263 //src_stride=(src_stride-width)*2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
264 height=(int)(&dither[height][0]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
265 asm volatile( |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
266 "movl %5, %%edx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
267 "movl %6, %%esi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
268 "movl %7, %%edi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
269 "movl %1, %%eax \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
270 "movd %%edx, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
271 "xorl $-1, %%edx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
272 "movl %%eax, %%ecx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
273 "addl $7, %%edx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
274 "subl %0, %%ecx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
275 "addl %%ecx, %%ecx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
276 "movd %%edx, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
277 "movl %%ecx, %1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
278 "movl %2, %%edx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
279 "shll $5, %%eax \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
280 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
281 "2: \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
282 "movq (%%edx), %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
283 "movq %%mm3, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
284 "pxor %%mm7, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
285 "punpcklbw %%mm7, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
286 "punpckhbw %%mm7, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
287 "movl %0, %%ecx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
288 "psraw %%mm5, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
289 "psraw %%mm5, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
290 "1: \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
291 "movq (%%esi), %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
292 "movq 8(%%esi), %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
293 "paddw %%mm3, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
294 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
295 "paddw (%%esi,%%eax,), %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
296 "paddw %%mm4, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
297 "movq 8(%%esi,%%eax,), %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
298 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
299 "movq %%mm7, (%%esi,%%eax,) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
300 "psraw %%mm2, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
301 "paddw %%mm6, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
302 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
303 "movq %%mm7, 8(%%esi,%%eax,) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
304 "psraw %%mm2, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
305 "packuswb %%mm1, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
306 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
307 "movq %%mm0, (%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
308 "addl $16, %%esi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
309 "addl $8, %%edi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
310 "subl $8, %%ecx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
311 "jg 1b \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
312 "addl %1, %%esi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
313 "addl $8, %%edx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
314 "addl %3, %%edi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
315 "cmpl %4, %%edx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
316 "jl 2b \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
317 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
318 : |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
319 : "m" (width), "m" (src_stride), "g" (od), "m" (dst_stride), "g" (height), |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
320 "m" (log2_scale), "m" (src), "m" (dst) //input |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
321 : "%eax", "%ecx", "%edx", "%edi", "%esi" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
322 ); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
323 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
324 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
325 static void mul_thrmat_mmx(struct vf_priv_s *p, int q) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
326 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
327 int adr=(int)(&p->threshold_mtx_noq[0]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
328 asm volatile( |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
329 "movd %0, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
330 "addl $8*8*2, %%edi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
331 "movq 0*8(%%esi), %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
332 "punpcklwd %%mm7, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
333 "movq 1*8(%%esi), %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
334 "punpckldq %%mm7, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
335 "pmullw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
336 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
337 "movq 2*8(%%esi), %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
338 "pmullw %%mm7, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
339 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
340 "movq 3*8(%%esi), %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
341 "pmullw %%mm7, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
342 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
343 "movq %%mm0, 0*8(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
344 "movq 4*8(%%esi), %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
345 "pmullw %%mm7, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
346 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
347 "movq %%mm1, 1*8(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
348 "movq 5*8(%%esi), %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
349 "pmullw %%mm7, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
350 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
351 "movq %%mm2, 2*8(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
352 "movq 6*8(%%esi), %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
353 "pmullw %%mm7, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
354 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
355 "movq %%mm3, 3*8(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
356 "movq 7*8+0*8(%%esi), %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
357 "pmullw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
358 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
359 "movq %%mm4, 4*8(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
360 "movq 7*8+1*8(%%esi), %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
361 "pmullw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
362 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
363 "movq %%mm5, 5*8(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
364 "movq 7*8+2*8(%%esi), %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
365 "pmullw %%mm7, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
366 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
367 "movq %%mm6, 6*8(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
368 "movq 7*8+3*8(%%esi), %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
369 "pmullw %%mm7, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
370 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
371 "movq %%mm0, 7*8+0*8(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
372 "movq 7*8+4*8(%%esi), %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
373 "pmullw %%mm7, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
374 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
375 "movq %%mm1, 7*8+1*8(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
376 "movq 7*8+5*8(%%esi), %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
377 "pmullw %%mm7, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
378 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
379 "movq %%mm2, 7*8+2*8(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
380 "movq 7*8+6*8(%%esi), %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
381 "pmullw %%mm7, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
382 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
383 "movq %%mm3, 7*8+3*8(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
384 "movq 14*8+0*8(%%esi), %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
385 "pmullw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
386 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
387 "movq %%mm4, 7*8+4*8(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
388 "movq 14*8+1*8(%%esi), %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
389 "pmullw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
390 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
391 "movq %%mm5, 7*8+5*8(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
392 "pmullw %%mm7, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
393 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
394 "movq %%mm6, 7*8+6*8(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
395 "movq %%mm0, 14*8+0*8(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
396 "movq %%mm1, 14*8+1*8(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
397 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
398 : "+g" (q), "+S" (adr), "+D" (adr) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
399 : |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
400 ); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
401 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
402 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
403 static void column_fidct_mmx(int16_t* thr_adr, DCTELEM *data, DCTELEM *output, int cnt); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
404 static void row_idct_mmx(DCTELEM* workspace, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
405 int16_t* output_adr, int output_stride, int cnt); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
406 static void row_fdct_mmx(DCTELEM *data, const uint8_t *pixels, int line_size, int cnt); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
407 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
408 #define store_slice_s store_slice_mmx |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
409 #define store_slice2_s store_slice2_mmx |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
410 #define mul_thrmat_s mul_thrmat_mmx |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
411 #define column_fidct_s column_fidct_mmx |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
412 #define row_idct_s row_idct_mmx |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
413 #define row_fdct_s row_fdct_mmx |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
414 #endif // HAVE_MMX |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
415 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
416 static void filter(struct vf_priv_s *p, uint8_t *dst, uint8_t *src, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
417 int dst_stride, int src_stride, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
418 int width, int height, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
419 uint8_t *qp_store, int qp_stride, int is_luma) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
420 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
421 int x, x0, y, es, qy, t; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
422 const int stride= is_luma ? p->temp_stride : (width+16);//((width+16+15)&(~15)) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
423 const int step=6-p->log2_count; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
424 const int qps= 3 + is_luma; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
425 int32_t block_align1[4*8*BLOCKSZ+ 4*8*BLOCKSZ+8];//32 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
426 int32_t *block_align=(int32_t*)(((int)block_align1+31)&(~31)); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
427 DCTELEM *block= (DCTELEM *)block_align; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
428 DCTELEM *block3=(DCTELEM *)(block_align+4*8*BLOCKSZ); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
429 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
430 memset(block3, 0, 4*8*BLOCKSZ); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
431 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
432 //p->src=src-src_stride*8-8;//! |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
433 if (!src || !dst) return; // HACK avoid crash for Y8 colourspace |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
434 for(y=0; y<height; y++){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
435 int index= 8 + 8*stride + y*stride; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
436 memcpy(p->src + index, src + y*src_stride, width);//this line can be avoided by using DR & user fr.buffers |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
437 for(x=0; x<8; x++){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
438 p->src[index - x - 1]= p->src[index + x ]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
439 p->src[index + width + x ]= p->src[index + width - x - 1]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
440 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
441 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
442 for(y=0; y<8; y++){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
443 memcpy(p->src + ( 7-y)*stride, p->src + ( y+8)*stride, stride); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
444 memcpy(p->src + (height+8+y)*stride, p->src + (height-y+7)*stride, stride); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
445 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
446 //FIXME (try edge emu) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
447 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
448 for(y=8; y<24; y++) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
449 memset(p->temp+ 8 +y*stride, 0,width*sizeof(int16_t)); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
450 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
451 for(y=step; y<height+8; y+=step){ //step= 1,2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
452 qy=y-4; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
453 if (qy>height-1) qy=height-1; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
454 if (qy<0) qy=0; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
455 qy=(qy>>qps)*qp_stride; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
456 row_fdct_s(block, p->src + y*stride +2-(y&1), stride, 2); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
457 for(x0=0; x0<width+8-8*(BLOCKSZ-1); x0+=8*(BLOCKSZ-1)){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
458 row_fdct_s(block+8*8, p->src + y*stride+8+x0 +2-(y&1), stride, 2*(BLOCKSZ-1)); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
459 if(p->qp) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
460 column_fidct_s((int16_t*)(&p->threshold_mtx[0]), block+0*8, block3+0*8, 8*(BLOCKSZ-1)); //yes, this is a HOTSPOT |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
461 else |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
462 for (x=0; x<8*(BLOCKSZ-1); x+=8) { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
463 t=x+x0-2; //correct t=x+x0-2-(y&1), but its the same |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
464 if (t<0) t=0;//t always < width-2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
465 t=qp_store[qy+(t>>qps)]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
466 if(p->mpeg2) t>>=1; //copy p->mpeg2,prev_q to locals? |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
467 if (t!=p->prev_q) p->prev_q=t, mul_thrmat_s(p, t); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
468 column_fidct_s((int16_t*)(&p->threshold_mtx[0]), block+x*8, block3+x*8, 8); //yes, this is a HOTSPOT |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
469 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
470 row_idct_s(block3+0*8, p->temp + (y&15)*stride+x0+2-(y&1), stride, 2*(BLOCKSZ-1)); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
471 memcpy(block, block+(BLOCKSZ-1)*64, 8*8*sizeof(DCTELEM)); //cycling |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
472 memcpy(block3, block3+(BLOCKSZ-1)*64, 6*8*sizeof(DCTELEM)); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
473 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
474 // |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
475 es=width+8-x0; // 8, ... |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
476 if (es>8) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
477 row_fdct_s(block+8*8, p->src + y*stride+8+x0 +2-(y&1), stride, (es-4)>>2); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
478 column_fidct_s((int16_t*)(&p->threshold_mtx[0]), block, block3, es-0); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
479 row_idct_s(block3+0*8, p->temp + (y&15)*stride+x0+2-(y&1), stride, es>>2); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
480 {const int y1=y-8+step;//l5-7 l4-6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
481 if (!(y1&7) && y1) { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
482 if (y1&8) store_slice_s(dst + (y1-8)*dst_stride, p->temp+ 8 +8*stride, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
483 dst_stride, stride, width, 8, 5-p->log2_count); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
484 else store_slice2_s(dst + (y1-8)*dst_stride, p->temp+ 8 +0*stride, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
485 dst_stride, stride, width, 8, 5-p->log2_count); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
486 } } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
487 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
488 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
489 if (y&7) { // == height & 7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
490 if (y&8) store_slice_s(dst + ((y-8)&~7)*dst_stride, p->temp+ 8 +8*stride, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
491 dst_stride, stride, width, y&7, 5-p->log2_count); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
492 else store_slice2_s(dst + ((y-8)&~7)*dst_stride, p->temp+ 8 +0*stride, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
493 dst_stride, stride, width, y&7, 5-p->log2_count); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
494 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
495 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
496 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
497 static int config(struct vf_instance_s* vf, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
498 int width, int height, int d_width, int d_height, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
499 unsigned int flags, unsigned int outfmt) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
500 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
501 int h= (height+16+15)&(~15); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
502 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
503 vf->priv->temp_stride= (width+16+15)&(~15); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
504 vf->priv->temp= (int16_t*)av_mallocz(vf->priv->temp_stride*3*8*sizeof(int16_t)); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
505 //this can also be avoided, see above |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
506 vf->priv->src = (uint8_t*)av_malloc(vf->priv->temp_stride*h*sizeof(uint8_t)); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
507 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
508 return vf_next_config(vf,width,height,d_width,d_height,flags,outfmt); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
509 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
510 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
511 static void get_image(struct vf_instance_s* vf, mp_image_t *mpi) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
512 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
513 if(mpi->flags&MP_IMGFLAG_PRESERVE) return; // don't change |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
514 // ok, we can do pp in-place (or pp disabled): |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
515 vf->dmpi=vf_get_image(vf->next,mpi->imgfmt, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
516 mpi->type, mpi->flags, mpi->w, mpi->h); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
517 mpi->planes[0]=vf->dmpi->planes[0]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
518 mpi->stride[0]=vf->dmpi->stride[0]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
519 mpi->width=vf->dmpi->width; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
520 if(mpi->flags&MP_IMGFLAG_PLANAR){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
521 mpi->planes[1]=vf->dmpi->planes[1]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
522 mpi->planes[2]=vf->dmpi->planes[2]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
523 mpi->stride[1]=vf->dmpi->stride[1]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
524 mpi->stride[2]=vf->dmpi->stride[2]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
525 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
526 mpi->flags|=MP_IMGFLAG_DIRECT; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
527 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
528 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
529 static int put_image(struct vf_instance_s* vf, mp_image_t *mpi) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
530 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
531 mp_image_t *dmpi; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
532 if(!(mpi->flags&MP_IMGFLAG_DIRECT)){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
533 // no DR, so get a new image! hope we'll get DR buffer: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
534 dmpi=vf_get_image(vf->next,mpi->imgfmt, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
535 MP_IMGTYPE_TEMP, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
536 MP_IMGFLAG_ACCEPT_STRIDE|MP_IMGFLAG_PREFER_ALIGNED_STRIDE, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
537 mpi->w,mpi->h); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
538 vf_clone_mpi_attributes(dmpi, mpi); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
539 }else{ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
540 dmpi=vf->dmpi; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
541 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
542 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
543 vf->priv->mpeg2= mpi->qscale_type; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
544 if(vf->priv->log2_count || !(mpi->flags&MP_IMGFLAG_DIRECT)){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
545 if(mpi->qscale || vf->priv->qp){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
546 filter(vf->priv, dmpi->planes[0], mpi->planes[0], dmpi->stride[0], mpi->stride[0], mpi->w, mpi->h, mpi->qscale, mpi->qstride, 1); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
547 filter(vf->priv, dmpi->planes[1], mpi->planes[1], dmpi->stride[1], mpi->stride[1], mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, mpi->qscale, mpi->qstride, 0); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
548 filter(vf->priv, dmpi->planes[2], mpi->planes[2], dmpi->stride[2], mpi->stride[2], mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, mpi->qscale, mpi->qstride, 0); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
549 }else{ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
550 memcpy_pic(dmpi->planes[0], mpi->planes[0], mpi->w, mpi->h, dmpi->stride[0], mpi->stride[0]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
551 memcpy_pic(dmpi->planes[1], mpi->planes[1], mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, dmpi->stride[1], mpi->stride[1]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
552 memcpy_pic(dmpi->planes[2], mpi->planes[2], mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, dmpi->stride[2], mpi->stride[2]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
553 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
554 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
555 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
556 #ifdef HAVE_MMX |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
557 if(gCpuCaps.hasMMX) asm volatile ("emms\n\t"); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
558 #endif |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
559 #ifdef HAVE_MMX2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
560 if(gCpuCaps.hasMMX2) asm volatile ("sfence\n\t"); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
561 #endif |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
562 return vf_next_put_image(vf,dmpi); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
563 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
564 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
565 static void uninit(struct vf_instance_s* vf) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
566 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
567 if(!vf->priv) return; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
568 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
569 if(vf->priv->temp) av_free(vf->priv->temp); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
570 vf->priv->temp= NULL; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
571 if(vf->priv->src) av_free(vf->priv->src); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
572 vf->priv->src= NULL; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
573 //if(vf->priv->avctx) free(vf->priv->avctx); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
574 //vf->priv->avctx= NULL; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
575 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
576 av_free(vf->priv); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
577 vf->priv=NULL; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
578 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
579 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
580 //===========================================================================// |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
581 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
582 static int query_format(struct vf_instance_s* vf, unsigned int fmt) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
583 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
584 switch(fmt){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
585 case IMGFMT_YVU9: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
586 case IMGFMT_IF09: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
587 case IMGFMT_YV12: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
588 case IMGFMT_I420: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
589 case IMGFMT_IYUV: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
590 case IMGFMT_CLPL: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
591 case IMGFMT_Y800: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
592 case IMGFMT_Y8: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
593 case IMGFMT_444P: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
594 case IMGFMT_422P: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
595 case IMGFMT_411P: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
596 return vf_next_query_format(vf,fmt); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
597 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
598 return 0; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
599 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
600 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
601 /* |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
602 static unsigned int fmt_list[]={ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
603 IMGFMT_YVU9, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
604 IMGFMT_IF09, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
605 IMGFMT_YV12, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
606 IMGFMT_I420, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
607 IMGFMT_IYUV, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
608 IMGFMT_CLPL, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
609 IMGFMT_Y800, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
610 IMGFMT_Y8, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
611 IMGFMT_444P, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
612 IMGFMT_422P, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
613 IMGFMT_411P, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
614 0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
615 }; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
616 */ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
617 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
618 static int control(struct vf_instance_s* vf, int request, void* data) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
619 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
620 switch(request){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
621 case VFCTRL_QUERY_MAX_PP_LEVEL: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
622 return 5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
623 case VFCTRL_SET_PP_LEVEL: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
624 vf->priv->log2_count= *((unsigned int*)data); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
625 if (vf->priv->log2_count < 4) vf->priv->log2_count=4; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
626 return CONTROL_TRUE; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
627 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
628 return vf_next_control(vf,request,data); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
629 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
630 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
631 static int open(vf_instance_t *vf, char* args) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
632 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
633 int i=0, bias; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
634 int custom_threshold_m[64]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
635 int log2c=-1; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
636 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
637 vf->config=config; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
638 vf->put_image=put_image; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
639 vf->get_image=get_image; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
640 vf->query_format=query_format; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
641 vf->uninit=uninit; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
642 vf->control= control; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
643 vf->priv=av_mallocz(sizeof(struct vf_priv_s));//assumes align 16 ! |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
644 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
645 avcodec_init(); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
646 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
647 //vf->priv->avctx= avcodec_alloc_context(); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
648 //dsputil_init(&vf->priv->dsp, vf->priv->avctx); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
649 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
650 vf->priv->log2_count= 4; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
651 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
652 if (args) sscanf(args, "%d:%d:%d", &log2c, &vf->priv->qp, &i); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
653 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
654 if( log2c >=4 && log2c <=5 ) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
655 vf->priv->log2_count = log2c; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
656 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
657 if(vf->priv->qp < 0) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
658 vf->priv->qp = 0; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
659 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
660 bias= (1<<4)+i; //regulable |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
661 vf->priv->prev_q=0; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
662 // |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
663 for(i=0;i<64;i++) //FIXME: tune custom_threshold[] and remove this ! |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
664 custom_threshold_m[i]=(int)(custom_threshold[i]*(bias/71.)+ 0.5); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
665 for(i=0;i<8;i++){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
666 vf->priv->threshold_mtx_noq[2*i]=(uint64_t)custom_threshold_m[i*8+2] |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
667 |(((uint64_t)custom_threshold_m[i*8+6])<<16) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
668 |(((uint64_t)custom_threshold_m[i*8+0])<<32) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
669 |(((uint64_t)custom_threshold_m[i*8+4])<<48); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
670 vf->priv->threshold_mtx_noq[2*i+1]=(uint64_t)custom_threshold_m[i*8+5] |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
671 |(((uint64_t)custom_threshold_m[i*8+3])<<16) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
672 |(((uint64_t)custom_threshold_m[i*8+1])<<32) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
673 |(((uint64_t)custom_threshold_m[i*8+7])<<48); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
674 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
675 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
676 if (vf->priv->qp) vf->priv->prev_q=vf->priv->qp, mul_thrmat_s(vf->priv, vf->priv->qp); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
677 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
678 return 1; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
679 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
680 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
681 vf_info_t vf_info_fspp = { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
682 "fast simple postprocess", |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
683 "fspp", |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
684 "Michael Niedermayer, Nikolaj Poroshin", |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
685 "", |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
686 open, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
687 NULL |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
688 }; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
689 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
690 //==================================================================== |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
691 //Specific spp's dct, idct and threshold functions |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
692 //I'd prefer to have them in the separate file. |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
693 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
694 #include "../mangle.h" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
695 //#define MANGLE(a) #a |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
696 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
697 //typedef int16_t DCTELEM; //! only int16_t |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
698 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
699 #define DCTSIZE 8 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
700 #define DCTSIZE_S "8" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
701 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
702 #define FIX(x,s) ((int) ((x) * (1<<s) + 0.5)&0xffff) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
703 #define C64(x) ((uint64_t)((x)|(x)<<16))<<32 | (uint64_t)(x) | (uint64_t)(x)<<16 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
704 #define FIX64(x,s) C64(FIX(x,s)) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
705 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
706 #define MULTIPLY16H(x,k) (((x)*(k))>>16) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
707 #define THRESHOLD(r,x,t) if(((unsigned)((x)+t))>t*2) r=(x);else r=0; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
708 #define DESCALE(x,n) (((x) + (1 << ((n)-1))) >> n) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
709 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
710 #ifdef HAVE_MMX |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
711 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
712 static uint64_t attribute_used __attribute__((aligned(8))) temps[4];//!! |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
713 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
714 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_382683433=FIX64(0.382683433, 14); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
715 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_541196100=FIX64(0.541196100, 14); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
716 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_707106781=FIX64(0.707106781, 14); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
717 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_1_306562965=FIX64(1.306562965, 14); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
718 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
719 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_1_414213562_A=FIX64(1.414213562, 14); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
720 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
721 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_1_847759065=FIX64(1.847759065, 13); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
722 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_2_613125930=FIX64(-2.613125930, 13); //- |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
723 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_1_414213562=FIX64(1.414213562, 13); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
724 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_1_082392200=FIX64(1.082392200, 13); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
725 //for t3,t5,t7 == 0 shortcut |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
726 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_847759065=FIX64(0.847759065, 14); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
727 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_566454497=FIX64(0.566454497, 14); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
728 static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_198912367=FIX64(0.198912367, 14); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
729 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
730 static uint64_t attribute_used __attribute__((aligned(8))) MM_DESCALE_RND=C64(4); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
731 static uint64_t attribute_used __attribute__((aligned(8))) MM_2=C64(2); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
732 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
733 #else /* !HAVE_MMX */ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
734 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
735 typedef int32_t int_simd16_t; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
736 static int16_t FIX_0_382683433=FIX(0.382683433, 14); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
737 static int16_t FIX_0_541196100=FIX(0.541196100, 14); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
738 static int16_t FIX_0_707106781=FIX(0.707106781, 14); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
739 static int16_t FIX_1_306562965=FIX(1.306562965, 14); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
740 static int16_t FIX_1_414213562_A=FIX(1.414213562, 14); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
741 static int16_t FIX_1_847759065=FIX(1.847759065, 13); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
742 static int16_t FIX_2_613125930=FIX(-2.613125930, 13); //- |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
743 static int16_t FIX_1_414213562=FIX(1.414213562, 13); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
744 static int16_t FIX_1_082392200=FIX(1.082392200, 13); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
745 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
746 #endif |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
747 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
748 #ifndef HAVE_MMX |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
749 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
750 static void column_fidct_c(int16_t* thr_adr, DCTELEM *data, DCTELEM *output, int cnt) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
751 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
752 int_simd16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
753 int_simd16_t tmp10, tmp11, tmp12, tmp13; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
754 int_simd16_t z1,z2,z3,z4,z5, z10, z11, z12, z13; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
755 int_simd16_t d0, d1, d2, d3, d4, d5, d6, d7; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
756 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
757 DCTELEM* dataptr; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
758 DCTELEM* wsptr; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
759 int16_t *threshold; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
760 int ctr; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
761 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
762 dataptr = data; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
763 wsptr = output; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
764 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
765 for (; cnt > 0; cnt-=2) { //start positions |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
766 threshold=(int16_t*)thr_adr;//threshold_mtx |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
767 for (ctr = DCTSIZE; ctr > 0; ctr--) { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
768 // Process columns from input, add to output. |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
769 tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
770 tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
771 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
772 tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
773 tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
774 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
775 tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
776 tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
777 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
778 tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
779 tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
780 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
781 // Even part of FDCT |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
782 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
783 tmp10 = tmp0 + tmp3; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
784 tmp13 = tmp0 - tmp3; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
785 tmp11 = tmp1 + tmp2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
786 tmp12 = tmp1 - tmp2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
787 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
788 d0 = tmp10 + tmp11; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
789 d4 = tmp10 - tmp11; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
790 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
791 z1 = MULTIPLY16H((tmp12 + tmp13) <<2, FIX_0_707106781); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
792 d2 = tmp13 + z1; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
793 d6 = tmp13 - z1; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
794 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
795 // Even part of IDCT |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
796 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
797 THRESHOLD(tmp0, d0, threshold[0*8]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
798 THRESHOLD(tmp1, d2, threshold[2*8]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
799 THRESHOLD(tmp2, d4, threshold[4*8]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
800 THRESHOLD(tmp3, d6, threshold[6*8]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
801 tmp0+=2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
802 tmp10 = (tmp0 + tmp2)>>2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
803 tmp11 = (tmp0 - tmp2)>>2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
804 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
805 tmp13 = (tmp1 + tmp3)>>2; //+2 ! (psnr decides) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
806 tmp12 = MULTIPLY16H((tmp1 - tmp3), FIX_1_414213562_A) - tmp13; //<<2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
807 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
808 tmp0 = tmp10 + tmp13; //->temps |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
809 tmp3 = tmp10 - tmp13; //->temps |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
810 tmp1 = tmp11 + tmp12; //->temps |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
811 tmp2 = tmp11 - tmp12; //->temps |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
812 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
813 // Odd part of FDCT |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
814 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
815 tmp10 = tmp4 + tmp5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
816 tmp11 = tmp5 + tmp6; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
817 tmp12 = tmp6 + tmp7; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
818 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
819 z5 = MULTIPLY16H((tmp10 - tmp12)<<2, FIX_0_382683433); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
820 z2 = MULTIPLY16H(tmp10 <<2, FIX_0_541196100) + z5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
821 z4 = MULTIPLY16H(tmp12 <<2, FIX_1_306562965) + z5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
822 z3 = MULTIPLY16H(tmp11 <<2, FIX_0_707106781); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
823 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
824 z11 = tmp7 + z3; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
825 z13 = tmp7 - z3; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
826 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
827 d5 = z13 + z2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
828 d3 = z13 - z2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
829 d1 = z11 + z4; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
830 d7 = z11 - z4; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
831 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
832 // Odd part of IDCT |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
833 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
834 THRESHOLD(tmp4, d1, threshold[1*8]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
835 THRESHOLD(tmp5, d3, threshold[3*8]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
836 THRESHOLD(tmp6, d5, threshold[5*8]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
837 THRESHOLD(tmp7, d7, threshold[7*8]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
838 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
839 //Simd version uses here a shortcut for the tmp5,tmp6,tmp7 == 0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
840 z13 = tmp6 + tmp5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
841 z10 = (tmp6 - tmp5)<<1; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
842 z11 = tmp4 + tmp7; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
843 z12 = (tmp4 - tmp7)<<1; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
844 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
845 tmp7 = (z11 + z13)>>2; //+2 ! |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
846 tmp11 = MULTIPLY16H((z11 - z13)<<1, FIX_1_414213562); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
847 z5 = MULTIPLY16H(z10 + z12, FIX_1_847759065); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
848 tmp10 = MULTIPLY16H(z12, FIX_1_082392200) - z5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
849 tmp12 = MULTIPLY16H(z10, FIX_2_613125930) + z5; // - !! |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
850 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
851 tmp6 = tmp12 - tmp7; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
852 tmp5 = tmp11 - tmp6; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
853 tmp4 = tmp10 + tmp5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
854 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
855 wsptr[DCTSIZE*0]+= (tmp0 + tmp7); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
856 wsptr[DCTSIZE*1]+= (tmp1 + tmp6); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
857 wsptr[DCTSIZE*2]+= (tmp2 + tmp5); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
858 wsptr[DCTSIZE*3]+= (tmp3 - tmp4); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
859 wsptr[DCTSIZE*4]+= (tmp3 + tmp4); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
860 wsptr[DCTSIZE*5]+= (tmp2 - tmp5); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
861 wsptr[DCTSIZE*6]= (tmp1 - tmp6); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
862 wsptr[DCTSIZE*7]= (tmp0 - tmp7); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
863 // |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
864 dataptr++; //next column |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
865 wsptr++; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
866 threshold++; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
867 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
868 dataptr+=8; //skip each second start pos |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
869 wsptr +=8; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
870 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
871 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
872 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
873 #else /* HAVE_MMX */ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
874 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
875 static void column_fidct_mmx(int16_t* thr_adr, DCTELEM *data, DCTELEM *output, int cnt) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
876 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
877 asm volatile( |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
878 ".align 16 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
879 "1: \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
880 "movq "DCTSIZE_S"*0*2(%%esi), %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
881 // |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
882 "movq "DCTSIZE_S"*3*2(%%esi), %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
883 "movq %%mm1, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
884 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
885 "paddw "DCTSIZE_S"*7*2(%%esi), %%mm1 \n\t" //t0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
886 "movq %%mm7, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
887 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
888 "paddw "DCTSIZE_S"*4*2(%%esi), %%mm7 \n\t" //t3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
889 "movq %%mm1, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
890 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
891 "movq "DCTSIZE_S"*1*2(%%esi), %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
892 "psubw %%mm7, %%mm1 \n\t" //t13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
893 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
894 "movq "DCTSIZE_S"*2*2(%%esi), %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
895 "movq %%mm6, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
896 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
897 "paddw "DCTSIZE_S"*6*2(%%esi), %%mm6 \n\t" //t1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
898 "paddw %%mm7, %%mm5 \n\t" //t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
899 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
900 "paddw "DCTSIZE_S"*5*2(%%esi), %%mm2 \n\t" //t2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
901 "movq %%mm6, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
902 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
903 "paddw %%mm2, %%mm6 \n\t" //t11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
904 "psubw %%mm2, %%mm7 \n\t" //t12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
905 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
906 "movq %%mm5, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
907 "paddw %%mm6, %%mm5 \n\t" //d0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
908 // i0 t13 t12 i3 i1 d0 - d4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
909 "psubw %%mm6, %%mm2 \n\t" //d4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
910 "paddw %%mm1, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
911 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
912 "movq 4*16(%%edx), %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
913 "psllw $2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
914 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
915 "psubw 0*16(%%edx), %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
916 "psubw %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
917 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
918 "paddusw 0*16(%%edx), %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
919 "paddusw %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
920 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
921 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
922 // |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
923 "paddw 0*16(%%edx), %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
924 "paddw %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
925 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
926 "psubusw 0*16(%%edx), %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
927 "psubusw %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
928 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
929 //This func is totally compute-bound, operates at huge speed. So, DC shortcut |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
930 // at this place isn't worthwhile due to BTB miss penalty (checked on Pent. 3). |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
931 //However, typical numbers: nondc - 29%%, dc - 46%%, zero - 25%%. All <> 0 case is very rare. |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
932 "paddw "MANGLE(MM_2)", %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
933 "movq %%mm2, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
934 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
935 "paddw %%mm5, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
936 "psubw %%mm6, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
937 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
938 "movq %%mm1, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
939 "paddw %%mm7, %%mm1 \n\t" //d2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
940 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
941 "psubw 2*16(%%edx), %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
942 "psubw %%mm7, %%mm6 \n\t" //d6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
943 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
944 "movq 6*16(%%edx), %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
945 "psraw $2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
946 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
947 "paddusw 2*16(%%edx), %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
948 "psubw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
949 // t7 d2 /t11 t4 t6 - d6 /t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
950 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
951 "paddw 2*16(%%edx), %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
952 "paddusw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
953 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
954 "psubusw 2*16(%%edx), %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
955 "paddw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
956 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
957 "psubw "DCTSIZE_S"*4*2(%%esi), %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
958 "psubusw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
959 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
960 //movq [edi+"DCTSIZE_S"*2*2], mm1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
961 //movq [edi+"DCTSIZE_S"*6*2], mm6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
962 "movq %%mm1, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
963 "psraw $2, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
964 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
965 "psubw "DCTSIZE_S"*6*2(%%esi), %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
966 "psubw %%mm6, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
967 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
968 "psubw "DCTSIZE_S"*7*2(%%esi), %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
969 "paddw %%mm7, %%mm6 \n\t" //'t13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
970 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
971 "psraw $2, %%mm6 \n\t" //paddw mm6, MM_2 !! --- |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
972 "movq %%mm2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
973 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
974 "pmulhw "MANGLE(MM_FIX_1_414213562_A)", %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
975 "paddw %%mm6, %%mm2 \n\t" //'t0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
976 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
977 "movq %%mm2, "MANGLE(temps)"+0*8 \n\t" //! |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
978 "psubw %%mm6, %%mm7 \n\t" //'t3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
979 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
980 "movq "DCTSIZE_S"*2*2(%%esi), %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
981 "psubw %%mm6, %%mm1 \n\t" //'t12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
982 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
983 "psubw "DCTSIZE_S"*5*2(%%esi), %%mm2 \n\t" //t5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
984 "movq %%mm5, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
985 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
986 "movq %%mm7, "MANGLE(temps)"+3*8 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
987 "paddw %%mm2, %%mm3 \n\t" //t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
988 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
989 "paddw %%mm4, %%mm2 \n\t" //t11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
990 "paddw %%mm0, %%mm4 \n\t" //t12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
991 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
992 "movq %%mm3, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
993 "psubw %%mm4, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
994 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
995 "psllw $2, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
996 "psllw $2, %%mm7 \n\t" //opt for P6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
997 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
998 "pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
999 "psllw $2, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1000 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1001 "pmulhw "MANGLE(MM_FIX_0_541196100)", %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1002 "psllw $2, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1003 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1004 "pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1005 "paddw %%mm1, %%mm5 \n\t" //'t1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1006 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1007 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1008 "psubw %%mm1, %%mm6 \n\t" //'t2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1009 // t7 't12 't11 t4 t6 - 't13 't10 --- |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1010 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1011 "paddw %%mm3, %%mm7 \n\t" //z2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1012 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1013 "movq %%mm5, "MANGLE(temps)"+1*8 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1014 "paddw %%mm3, %%mm4 \n\t" //z4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1015 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1016 "movq 3*16(%%edx), %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1017 "movq %%mm0, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1018 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1019 "movq %%mm6, "MANGLE(temps)"+2*8 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1020 "psubw %%mm2, %%mm1 \n\t" //z13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1021 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1022 //=== |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1023 "paddw %%mm2, %%mm0 \n\t" //z11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1024 "movq %%mm1, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1025 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1026 "movq 5*16(%%edx), %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1027 "psubw %%mm7, %%mm1 \n\t" //d3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1028 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1029 "paddw %%mm7, %%mm5 \n\t" //d5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1030 "psubw %%mm3, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1031 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1032 "movq 1*16(%%edx), %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1033 "psubw %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1034 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1035 "movq %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1036 "paddw %%mm4, %%mm0 \n\t" //d1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1037 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1038 "paddusw %%mm3, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1039 "psubw %%mm4, %%mm6 \n\t" //d7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1040 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1041 // d1 d3 - - - d5 d7 - |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1042 "movq 7*16(%%edx), %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1043 "psubw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1044 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1045 "psubw %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1046 "paddusw %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1047 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1048 "paddusw %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1049 "paddw %%mm3, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1050 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1051 "paddw %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1052 "paddw %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1053 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1054 "psubusw %%mm3, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1055 "psubusw %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1056 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1057 "psubusw %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1058 "movq %%mm1, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1059 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1060 "por %%mm5, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1061 "paddusw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1062 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1063 "por %%mm6, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1064 "paddw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1065 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1066 "packssdw %%mm4, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1067 "psubusw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1068 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1069 "movd %%mm4, %%eax \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1070 "orl %%eax, %%eax \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1071 "jnz 2f \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1072 //movq [edi+"DCTSIZE_S"*3*2], mm1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1073 //movq [edi+"DCTSIZE_S"*5*2], mm5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1074 //movq [edi+"DCTSIZE_S"*1*2], mm0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1075 //movq [edi+"DCTSIZE_S"*7*2], mm6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1076 // t4 t5 - - - t6 t7 - |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1077 //--- t4 (mm0) may be <>0; mm1, mm5, mm6 == 0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1078 //Typical numbers: nondc - 19%%, dc - 26%%, zero - 55%%. zero case alone isn't worthwhile |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1079 "movq "MANGLE(temps)"+0*8, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1080 "movq %%mm0, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1081 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1082 "pmulhw "MANGLE(MM_FIX_0_847759065)", %%mm0 \n\t" //tmp6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1083 "movq %%mm1, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1084 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1085 "movq "DCTSIZE_S"*0*2(%%edi), %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1086 "movq %%mm2, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1087 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1088 "pmulhw "MANGLE(MM_FIX_0_566454497)", %%mm1 \n\t" //tmp5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1089 "paddw %%mm4, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1090 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1091 "movq "MANGLE(temps)"+1*8, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1092 //paddw mm3, MM_2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1093 "psraw $2, %%mm3 \n\t" //tmp7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1094 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1095 "pmulhw "MANGLE(MM_FIX_0_198912367)", %%mm2 \n\t" //-tmp4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1096 "psubw %%mm3, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1097 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1098 "movq "DCTSIZE_S"*1*2(%%edi), %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1099 "paddw %%mm3, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1100 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1101 "movq %%mm4, "DCTSIZE_S"*7*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1102 "paddw %%mm6, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1103 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1104 "movq "MANGLE(temps)"+2*8, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1105 "psubw %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1106 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1107 "movq "DCTSIZE_S"*2*2(%%edi), %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1108 "paddw %%mm0, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1109 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1110 "movq %%mm5, "DCTSIZE_S"*0*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1111 "paddw %%mm3, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1112 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1113 "movq %%mm6, "DCTSIZE_S"*6*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1114 "psubw %%mm1, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1115 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1116 "movq "DCTSIZE_S"*5*2(%%edi), %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1117 "paddw %%mm1, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1118 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1119 "movq "DCTSIZE_S"*3*2(%%edi), %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1120 "paddw %%mm3, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1121 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1122 "movq "MANGLE(temps)"+3*8, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1123 "addl $8, %%esi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1124 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1125 "movq %%mm7, "DCTSIZE_S"*1*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1126 "paddw %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1127 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1128 "movq %%mm4, "DCTSIZE_S"*2*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1129 "psubw %%mm2, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1130 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1131 "movq "DCTSIZE_S"*4*2(%%edi), %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1132 "paddw %%mm2, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1133 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1134 "movq %%mm5, "DCTSIZE_S"*5*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1135 "paddw %%mm0, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1136 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1137 "movq %%mm6, "DCTSIZE_S"*3*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1138 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1139 "movq %%mm7, "DCTSIZE_S"*4*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1140 "addl $8, %%edi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1141 "jmp 4f \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1142 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1143 "2: \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1144 //--- non DC2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1145 //psraw mm1, 2 w/o it -> offset. thr1, thr1, thr1 (actually thr1, thr1, thr1-1) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1146 //psraw mm5, 2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1147 //psraw mm0, 2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1148 //psraw mm6, 2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1149 "movq %%mm5, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1150 "psubw %%mm1, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1151 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1152 "psllw $1, %%mm5 \n\t" //'z10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1153 "paddw %%mm1, %%mm3 \n\t" //'z13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1154 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1155 "movq %%mm0, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1156 "psubw %%mm6, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1157 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1158 "movq %%mm5, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1159 "psllw $1, %%mm0 \n\t" //'z12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1160 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1161 "pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm1 \n\t" //- |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1162 "paddw %%mm0, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1163 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1164 "pmulhw "MANGLE(MM_FIX_1_847759065)", %%mm5 \n\t" //'z5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1165 "paddw %%mm6, %%mm2 \n\t" //'z11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1166 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1167 "pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1168 "movq %%mm2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1169 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1170 //--- |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1171 "movq "MANGLE(temps)"+0*8, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1172 "psubw %%mm3, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1173 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1174 "psllw $1, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1175 "paddw %%mm3, %%mm7 \n\t" //'t7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1176 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1177 "pmulhw "MANGLE(MM_FIX_1_414213562)", %%mm2 \n\t" //'t11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1178 "movq %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1179 //paddw mm7, MM_2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1180 "psraw $2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1181 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1182 "paddw "DCTSIZE_S"*0*2(%%edi), %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1183 "psubw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1184 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1185 "movq "MANGLE(temps)"+1*8, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1186 "paddw %%mm7, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1187 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1188 "movq %%mm6, "DCTSIZE_S"*7*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1189 "paddw %%mm5, %%mm1 \n\t" //'t12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1190 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1191 "movq %%mm4, "DCTSIZE_S"*0*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1192 "psubw %%mm7, %%mm1 \n\t" //'t6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1193 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1194 "movq "MANGLE(temps)"+2*8, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1195 "psubw %%mm5, %%mm0 \n\t" //'t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1196 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1197 "movq "MANGLE(temps)"+3*8, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1198 "movq %%mm3, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1199 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1200 "paddw "DCTSIZE_S"*1*2(%%edi), %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1201 "psubw %%mm1, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1202 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1203 "psubw %%mm1, %%mm2 \n\t" //'t5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1204 "paddw %%mm1, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1205 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1206 "movq %%mm5, "DCTSIZE_S"*6*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1207 "movq %%mm7, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1208 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1209 "paddw "DCTSIZE_S"*2*2(%%edi), %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1210 "psubw %%mm2, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1211 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1212 "paddw "DCTSIZE_S"*5*2(%%edi), %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1213 "paddw %%mm2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1214 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1215 "movq %%mm3, "DCTSIZE_S"*1*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1216 "paddw %%mm2, %%mm0 \n\t" //'t4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1217 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1218 // 't4 't6 't5 - - - - 't7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1219 "movq %%mm7, "DCTSIZE_S"*2*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1220 "movq %%mm6, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1221 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1222 "paddw "DCTSIZE_S"*4*2(%%edi), %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1223 "psubw %%mm0, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1224 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1225 "paddw "DCTSIZE_S"*3*2(%%edi), %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1226 "paddw %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1227 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1228 "movq %%mm4, "DCTSIZE_S"*5*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1229 "addl $8, %%esi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1230 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1231 "movq %%mm6, "DCTSIZE_S"*4*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1232 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1233 "movq %%mm1, "DCTSIZE_S"*3*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1234 "addl $8, %%edi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1235 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1236 "4: \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1237 //=part 2 (the same)=========================================================== |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1238 "movq "DCTSIZE_S"*0*2(%%esi), %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1239 // |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1240 "movq "DCTSIZE_S"*3*2(%%esi), %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1241 "movq %%mm1, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1242 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1243 "paddw "DCTSIZE_S"*7*2(%%esi), %%mm1 \n\t" //t0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1244 "movq %%mm7, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1245 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1246 "paddw "DCTSIZE_S"*4*2(%%esi), %%mm7 \n\t" //t3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1247 "movq %%mm1, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1248 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1249 "movq "DCTSIZE_S"*1*2(%%esi), %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1250 "psubw %%mm7, %%mm1 \n\t" //t13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1251 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1252 "movq "DCTSIZE_S"*2*2(%%esi), %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1253 "movq %%mm6, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1254 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1255 "paddw "DCTSIZE_S"*6*2(%%esi), %%mm6 \n\t" //t1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1256 "paddw %%mm7, %%mm5 \n\t" //t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1257 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1258 "paddw "DCTSIZE_S"*5*2(%%esi), %%mm2 \n\t" //t2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1259 "movq %%mm6, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1260 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1261 "paddw %%mm2, %%mm6 \n\t" //t11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1262 "psubw %%mm2, %%mm7 \n\t" //t12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1263 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1264 "movq %%mm5, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1265 "paddw %%mm6, %%mm5 \n\t" //d0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1266 // i0 t13 t12 i3 i1 d0 - d4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1267 "psubw %%mm6, %%mm2 \n\t" //d4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1268 "paddw %%mm1, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1269 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1270 "movq 1*8+4*16(%%edx), %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1271 "psllw $2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1272 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1273 "psubw 1*8+0*16(%%edx), %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1274 "psubw %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1275 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1276 "paddusw 1*8+0*16(%%edx), %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1277 "paddusw %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1278 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1279 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1280 // |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1281 "paddw 1*8+0*16(%%edx), %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1282 "paddw %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1283 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1284 "psubusw 1*8+0*16(%%edx), %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1285 "psubusw %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1286 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1287 //This func is totally compute-bound, operates at huge speed. So, DC shortcut |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1288 // at this place isn't worthwhile due to BTB miss penalty (checked on Pent. 3). |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1289 //However, typical numbers: nondc - 29%%, dc - 46%%, zero - 25%%. All <> 0 case is very rare. |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1290 "paddw "MANGLE(MM_2)", %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1291 "movq %%mm2, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1292 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1293 "paddw %%mm5, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1294 "psubw %%mm6, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1295 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1296 "movq %%mm1, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1297 "paddw %%mm7, %%mm1 \n\t" //d2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1298 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1299 "psubw 1*8+2*16(%%edx), %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1300 "psubw %%mm7, %%mm6 \n\t" //d6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1301 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1302 "movq 1*8+6*16(%%edx), %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1303 "psraw $2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1304 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1305 "paddusw 1*8+2*16(%%edx), %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1306 "psubw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1307 // t7 d2 /t11 t4 t6 - d6 /t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1308 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1309 "paddw 1*8+2*16(%%edx), %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1310 "paddusw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1311 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1312 "psubusw 1*8+2*16(%%edx), %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1313 "paddw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1314 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1315 "psubw "DCTSIZE_S"*4*2(%%esi), %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1316 "psubusw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1317 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1318 //movq [edi+"DCTSIZE_S"*2*2], mm1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1319 //movq [edi+"DCTSIZE_S"*6*2], mm6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1320 "movq %%mm1, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1321 "psraw $2, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1322 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1323 "psubw "DCTSIZE_S"*6*2(%%esi), %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1324 "psubw %%mm6, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1325 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1326 "psubw "DCTSIZE_S"*7*2(%%esi), %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1327 "paddw %%mm7, %%mm6 \n\t" //'t13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1328 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1329 "psraw $2, %%mm6 \n\t" //paddw mm6, MM_2 !! --- |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1330 "movq %%mm2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1331 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1332 "pmulhw "MANGLE(MM_FIX_1_414213562_A)", %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1333 "paddw %%mm6, %%mm2 \n\t" //'t0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1334 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1335 "movq %%mm2, "MANGLE(temps)"+0*8 \n\t" //! |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1336 "psubw %%mm6, %%mm7 \n\t" //'t3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1337 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1338 "movq "DCTSIZE_S"*2*2(%%esi), %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1339 "psubw %%mm6, %%mm1 \n\t" //'t12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1340 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1341 "psubw "DCTSIZE_S"*5*2(%%esi), %%mm2 \n\t" //t5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1342 "movq %%mm5, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1343 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1344 "movq %%mm7, "MANGLE(temps)"+3*8 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1345 "paddw %%mm2, %%mm3 \n\t" //t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1346 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1347 "paddw %%mm4, %%mm2 \n\t" //t11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1348 "paddw %%mm0, %%mm4 \n\t" //t12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1349 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1350 "movq %%mm3, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1351 "psubw %%mm4, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1352 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1353 "psllw $2, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1354 "psllw $2, %%mm7 \n\t" //opt for P6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1355 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1356 "pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1357 "psllw $2, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1358 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1359 "pmulhw "MANGLE(MM_FIX_0_541196100)", %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1360 "psllw $2, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1361 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1362 "pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1363 "paddw %%mm1, %%mm5 \n\t" //'t1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1364 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1365 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1366 "psubw %%mm1, %%mm6 \n\t" //'t2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1367 // t7 't12 't11 t4 t6 - 't13 't10 --- |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1368 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1369 "paddw %%mm3, %%mm7 \n\t" //z2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1370 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1371 "movq %%mm5, "MANGLE(temps)"+1*8 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1372 "paddw %%mm3, %%mm4 \n\t" //z4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1373 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1374 "movq 1*8+3*16(%%edx), %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1375 "movq %%mm0, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1376 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1377 "movq %%mm6, "MANGLE(temps)"+2*8 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1378 "psubw %%mm2, %%mm1 \n\t" //z13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1379 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1380 //=== |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1381 "paddw %%mm2, %%mm0 \n\t" //z11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1382 "movq %%mm1, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1383 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1384 "movq 1*8+5*16(%%edx), %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1385 "psubw %%mm7, %%mm1 \n\t" //d3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1386 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1387 "paddw %%mm7, %%mm5 \n\t" //d5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1388 "psubw %%mm3, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1389 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1390 "movq 1*8+1*16(%%edx), %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1391 "psubw %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1392 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1393 "movq %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1394 "paddw %%mm4, %%mm0 \n\t" //d1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1395 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1396 "paddusw %%mm3, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1397 "psubw %%mm4, %%mm6 \n\t" //d7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1398 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1399 // d1 d3 - - - d5 d7 - |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1400 "movq 1*8+7*16(%%edx), %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1401 "psubw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1402 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1403 "psubw %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1404 "paddusw %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1405 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1406 "paddusw %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1407 "paddw %%mm3, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1408 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1409 "paddw %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1410 "paddw %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1411 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1412 "psubusw %%mm3, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1413 "psubusw %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1414 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1415 "psubusw %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1416 "movq %%mm1, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1417 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1418 "por %%mm5, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1419 "paddusw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1420 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1421 "por %%mm6, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1422 "paddw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1423 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1424 "packssdw %%mm4, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1425 "psubusw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1426 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1427 "movd %%mm4, %%eax \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1428 "orl %%eax, %%eax \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1429 "jnz 3f \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1430 //movq [edi+"DCTSIZE_S"*3*2], mm1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1431 //movq [edi+"DCTSIZE_S"*5*2], mm5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1432 //movq [edi+"DCTSIZE_S"*1*2], mm0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1433 //movq [edi+"DCTSIZE_S"*7*2], mm6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1434 // t4 t5 - - - t6 t7 - |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1435 //--- t4 (mm0) may be <>0; mm1, mm5, mm6 == 0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1436 //Typical numbers: nondc - 19%%, dc - 26%%, zero - 55%%. zero case alone isn't worthwhile |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1437 "movq "MANGLE(temps)"+0*8, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1438 "movq %%mm0, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1439 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1440 "pmulhw "MANGLE(MM_FIX_0_847759065)", %%mm0 \n\t" //tmp6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1441 "movq %%mm1, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1442 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1443 "movq "DCTSIZE_S"*0*2(%%edi), %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1444 "movq %%mm2, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1445 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1446 "pmulhw "MANGLE(MM_FIX_0_566454497)", %%mm1 \n\t" //tmp5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1447 "paddw %%mm4, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1448 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1449 "movq "MANGLE(temps)"+1*8, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1450 //paddw mm3, MM_2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1451 "psraw $2, %%mm3 \n\t" //tmp7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1452 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1453 "pmulhw "MANGLE(MM_FIX_0_198912367)", %%mm2 \n\t" //-tmp4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1454 "psubw %%mm3, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1455 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1456 "movq "DCTSIZE_S"*1*2(%%edi), %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1457 "paddw %%mm3, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1458 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1459 "movq %%mm4, "DCTSIZE_S"*7*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1460 "paddw %%mm6, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1461 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1462 "movq "MANGLE(temps)"+2*8, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1463 "psubw %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1464 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1465 "movq "DCTSIZE_S"*2*2(%%edi), %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1466 "paddw %%mm0, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1467 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1468 "movq %%mm5, "DCTSIZE_S"*0*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1469 "paddw %%mm3, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1470 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1471 "movq %%mm6, "DCTSIZE_S"*6*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1472 "psubw %%mm1, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1473 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1474 "movq "DCTSIZE_S"*5*2(%%edi), %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1475 "paddw %%mm1, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1476 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1477 "movq "DCTSIZE_S"*3*2(%%edi), %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1478 "paddw %%mm3, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1479 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1480 "movq "MANGLE(temps)"+3*8, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1481 "addl $24, %%esi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1482 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1483 "movq %%mm7, "DCTSIZE_S"*1*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1484 "paddw %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1485 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1486 "movq %%mm4, "DCTSIZE_S"*2*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1487 "psubw %%mm2, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1488 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1489 "movq "DCTSIZE_S"*4*2(%%edi), %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1490 "paddw %%mm2, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1491 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1492 "movq %%mm5, "DCTSIZE_S"*5*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1493 "paddw %%mm0, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1494 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1495 "movq %%mm6, "DCTSIZE_S"*3*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1496 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1497 "movq %%mm7, "DCTSIZE_S"*4*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1498 "addl $24, %%edi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1499 "subl $2, %%ecx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1500 "jnz 1b \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1501 "jmp 5f \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1502 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1503 "3: \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1504 //--- non DC2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1505 //psraw mm1, 2 w/o it -> offset. thr1, thr1, thr1 (actually thr1, thr1, thr1-1) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1506 //psraw mm5, 2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1507 //psraw mm0, 2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1508 //psraw mm6, 2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1509 "movq %%mm5, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1510 "psubw %%mm1, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1511 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1512 "psllw $1, %%mm5 \n\t" //'z10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1513 "paddw %%mm1, %%mm3 \n\t" //'z13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1514 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1515 "movq %%mm0, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1516 "psubw %%mm6, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1517 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1518 "movq %%mm5, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1519 "psllw $1, %%mm0 \n\t" //'z12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1520 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1521 "pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm1 \n\t" //- |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1522 "paddw %%mm0, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1523 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1524 "pmulhw "MANGLE(MM_FIX_1_847759065)", %%mm5 \n\t" //'z5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1525 "paddw %%mm6, %%mm2 \n\t" //'z11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1526 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1527 "pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1528 "movq %%mm2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1529 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1530 //--- |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1531 "movq "MANGLE(temps)"+0*8, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1532 "psubw %%mm3, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1533 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1534 "psllw $1, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1535 "paddw %%mm3, %%mm7 \n\t" //'t7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1536 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1537 "pmulhw "MANGLE(MM_FIX_1_414213562)", %%mm2 \n\t" //'t11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1538 "movq %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1539 //paddw mm7, MM_2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1540 "psraw $2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1541 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1542 "paddw "DCTSIZE_S"*0*2(%%edi), %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1543 "psubw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1544 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1545 "movq "MANGLE(temps)"+1*8, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1546 "paddw %%mm7, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1547 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1548 "movq %%mm6, "DCTSIZE_S"*7*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1549 "paddw %%mm5, %%mm1 \n\t" //'t12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1550 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1551 "movq %%mm4, "DCTSIZE_S"*0*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1552 "psubw %%mm7, %%mm1 \n\t" //'t6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1553 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1554 "movq "MANGLE(temps)"+2*8, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1555 "psubw %%mm5, %%mm0 \n\t" //'t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1556 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1557 "movq "MANGLE(temps)"+3*8, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1558 "movq %%mm3, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1559 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1560 "paddw "DCTSIZE_S"*1*2(%%edi), %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1561 "psubw %%mm1, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1562 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1563 "psubw %%mm1, %%mm2 \n\t" //'t5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1564 "paddw %%mm1, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1565 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1566 "movq %%mm5, "DCTSIZE_S"*6*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1567 "movq %%mm7, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1568 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1569 "paddw "DCTSIZE_S"*2*2(%%edi), %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1570 "psubw %%mm2, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1571 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1572 "paddw "DCTSIZE_S"*5*2(%%edi), %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1573 "paddw %%mm2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1574 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1575 "movq %%mm3, "DCTSIZE_S"*1*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1576 "paddw %%mm2, %%mm0 \n\t" //'t4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1577 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1578 // 't4 't6 't5 - - - - 't7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1579 "movq %%mm7, "DCTSIZE_S"*2*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1580 "movq %%mm6, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1581 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1582 "paddw "DCTSIZE_S"*4*2(%%edi), %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1583 "psubw %%mm0, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1584 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1585 "paddw "DCTSIZE_S"*3*2(%%edi), %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1586 "paddw %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1587 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1588 "movq %%mm4, "DCTSIZE_S"*5*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1589 "addl $24, %%esi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1590 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1591 "movq %%mm6, "DCTSIZE_S"*4*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1592 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1593 "movq %%mm1, "DCTSIZE_S"*3*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1594 "addl $24, %%edi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1595 "subl $2, %%ecx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1596 "jnz 1b \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1597 "5: \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1598 |
15632 | 1599 : "+S"(data), "+D"(output), "+c"(cnt)// input regs |
1600 : "d"(thr_adr) | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1601 : "%eax" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1602 ); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1603 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1604 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1605 #endif // HAVE_MMX |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1606 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1607 #ifndef HAVE_MMX |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1608 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1609 static void row_idct_c(DCTELEM* workspace, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1610 int16_t* output_adr, int output_stride, int cnt) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1611 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1612 int_simd16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1613 int_simd16_t tmp10, tmp11, tmp12, tmp13; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1614 int_simd16_t z5, z10, z11, z12, z13; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1615 int16_t* outptr; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1616 DCTELEM* wsptr; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1617 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1618 cnt*=4; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1619 wsptr = workspace; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1620 outptr = output_adr; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1621 for (; cnt > 0; cnt--) { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1622 // Even part |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1623 //Simd version reads 4x4 block and transposes it |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1624 tmp10 = ( wsptr[2] + wsptr[3]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1625 tmp11 = ( wsptr[2] - wsptr[3]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1626 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1627 tmp13 = ( wsptr[0] + wsptr[1]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1628 tmp12 = (MULTIPLY16H( wsptr[0] - wsptr[1], FIX_1_414213562_A)<<2) - tmp13;//this shift order to avoid overflow |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1629 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1630 tmp0 = tmp10 + tmp13; //->temps |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1631 tmp3 = tmp10 - tmp13; //->temps |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1632 tmp1 = tmp11 + tmp12; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1633 tmp2 = tmp11 - tmp12; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1634 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1635 // Odd part |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1636 //Also transpose, with previous: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1637 // ---- ---- |||| |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1638 // ---- ---- idct |||| |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1639 // ---- ---- ---> |||| |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1640 // ---- ---- |||| |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1641 z13 = wsptr[4] + wsptr[5]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1642 z10 = wsptr[4] - wsptr[5]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1643 z11 = wsptr[6] + wsptr[7]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1644 z12 = wsptr[6] - wsptr[7]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1645 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1646 tmp7 = z11 + z13; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1647 tmp11 = MULTIPLY16H(z11 - z13, FIX_1_414213562); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1648 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1649 z5 = MULTIPLY16H(z10 + z12, FIX_1_847759065); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1650 tmp10 = MULTIPLY16H(z12, FIX_1_082392200) - z5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1651 tmp12 = MULTIPLY16H(z10, FIX_2_613125930) + z5; // - FIX_ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1652 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1653 tmp6 = (tmp12<<3) - tmp7; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1654 tmp5 = (tmp11<<3) - tmp6; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1655 tmp4 = (tmp10<<3) + tmp5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1656 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1657 // Final output stage: descale and write column |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1658 outptr[0*output_stride]+= DESCALE(tmp0 + tmp7, 3); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1659 outptr[1*output_stride]+= DESCALE(tmp1 + tmp6, 3); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1660 outptr[2*output_stride]+= DESCALE(tmp2 + tmp5, 3); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1661 outptr[3*output_stride]+= DESCALE(tmp3 - tmp4, 3); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1662 outptr[4*output_stride]+= DESCALE(tmp3 + tmp4, 3); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1663 outptr[5*output_stride]+= DESCALE(tmp2 - tmp5, 3); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1664 outptr[6*output_stride]+= DESCALE(tmp1 - tmp6, 3); //no += ? |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1665 outptr[7*output_stride]+= DESCALE(tmp0 - tmp7, 3); //no += ? |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1666 outptr++; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1667 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1668 wsptr += DCTSIZE; // advance pointer to next row |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1669 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1670 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1671 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1672 #else /* HAVE_MMX */ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1673 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1674 static void row_idct_mmx (DCTELEM* workspace, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1675 int16_t* output_adr, int output_stride, int cnt) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1676 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1677 asm volatile( |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1678 "leal (%%eax,%%eax,2), %%edx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1679 "1: \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1680 "movq "DCTSIZE_S"*0*2(%%esi), %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1681 // |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1682 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1683 "movq "DCTSIZE_S"*1*2(%%esi), %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1684 "movq %%mm0, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1685 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1686 "movq "DCTSIZE_S"*2*2(%%esi), %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1687 "punpcklwd %%mm1, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1688 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1689 "movq "DCTSIZE_S"*3*2(%%esi), %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1690 "punpckhwd %%mm1, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1691 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1692 //transpose 4x4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1693 "movq %%mm2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1694 "punpcklwd %%mm3, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1695 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1696 "movq %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1697 "punpckldq %%mm2, %%mm0 \n\t" //0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1698 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1699 "punpckhdq %%mm2, %%mm6 \n\t" //1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1700 "movq %%mm0, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1701 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1702 "punpckhwd %%mm3, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1703 "psubw %%mm6, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1704 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1705 "pmulhw "MANGLE(MM_FIX_1_414213562_A)", %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1706 "movq %%mm4, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1707 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1708 "punpckldq %%mm7, %%mm4 \n\t" //2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1709 "paddw %%mm6, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1710 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1711 "punpckhdq %%mm7, %%mm2 \n\t" //3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1712 "movq %%mm4, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1713 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1714 "psllw $2, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1715 "paddw %%mm2, %%mm4 \n\t" //t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1716 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1717 "movq "DCTSIZE_S"*0*2+"DCTSIZE_S"(%%esi), %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1718 "psubw %%mm2, %%mm1 \n\t" //t11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1719 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1720 "movq "DCTSIZE_S"*1*2+"DCTSIZE_S"(%%esi), %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1721 "psubw %%mm5, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1722 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1723 "movq %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1724 "paddw %%mm5, %%mm4 \n\t" //t0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1725 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1726 "psubw %%mm5, %%mm6 \n\t" //t3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1727 "movq %%mm1, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1728 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1729 "movq "DCTSIZE_S"*2*2+"DCTSIZE_S"(%%esi), %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1730 "paddw %%mm0, %%mm1 \n\t" //t1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1731 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1732 "movq %%mm4, "MANGLE(temps)"+0*8 \n\t" //t0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1733 "movq %%mm3, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1734 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1735 "movq %%mm6, "MANGLE(temps)"+1*8 \n\t" //t3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1736 "punpcklwd %%mm2, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1737 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1738 //transpose 4x4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1739 "movq "DCTSIZE_S"*3*2+"DCTSIZE_S"(%%esi), %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1740 "punpckhwd %%mm2, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1741 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1742 "movq %%mm5, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1743 "punpcklwd %%mm6, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1744 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1745 "psubw %%mm0, %%mm7 \n\t" //t2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1746 "punpckhwd %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1747 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1748 "movq %%mm3, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1749 "punpckldq %%mm5, %%mm3 \n\t" //4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1750 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1751 "punpckhdq %%mm5, %%mm0 \n\t" //5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1752 "movq %%mm4, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1753 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1754 // |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1755 "movq %%mm3, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1756 "punpckldq %%mm2, %%mm4 \n\t" //6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1757 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1758 "psubw %%mm0, %%mm3 \n\t" //z10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1759 "punpckhdq %%mm2, %%mm5 \n\t" //7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1760 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1761 "paddw %%mm0, %%mm6 \n\t" //z13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1762 "movq %%mm4, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1763 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1764 "movq %%mm3, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1765 "psubw %%mm5, %%mm4 \n\t" //z12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1766 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1767 "pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm0 \n\t" //- |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1768 "paddw %%mm4, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1769 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1770 "pmulhw "MANGLE(MM_FIX_1_847759065)", %%mm3 \n\t" //z5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1771 "paddw %%mm5, %%mm2 \n\t" //z11 > |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1772 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1773 "pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1774 "movq %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1775 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1776 "psubw %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1777 "paddw %%mm6, %%mm5 \n\t" //t7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1778 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1779 "pmulhw "MANGLE(MM_FIX_1_414213562)", %%mm2 \n\t" //t11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1780 "paddw %%mm3, %%mm0 \n\t" //t12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1781 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1782 "psllw $3, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1783 "psubw %%mm3, %%mm4 \n\t" //t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1784 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1785 "movq "MANGLE(temps)"+0*8, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1786 "movq %%mm1, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1787 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1788 "psllw $3, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1789 "psubw %%mm5, %%mm0 \n\t" //t6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1790 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1791 "psllw $3, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1792 "paddw %%mm0, %%mm1 \n\t" //d1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1793 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1794 "psubw %%mm0, %%mm2 \n\t" //t5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1795 "psubw %%mm0, %%mm3 \n\t" //d6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1796 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1797 "paddw %%mm2, %%mm4 \n\t" //t4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1798 "movq %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1799 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1800 "paddw %%mm2, %%mm7 \n\t" //d2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1801 "psubw %%mm2, %%mm0 \n\t" //d5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1802 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1803 "movq "MANGLE(MM_DESCALE_RND)", %%mm2 \n\t" //4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1804 "psubw %%mm5, %%mm6 \n\t" //d7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1805 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1806 "paddw "MANGLE(temps)"+0*8, %%mm5 \n\t" //d0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1807 "paddw %%mm2, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1808 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1809 "paddw %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1810 "psraw $3, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1811 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1812 "paddw %%mm2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1813 "psraw $3, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1814 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1815 "paddw (%%edi), %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1816 "psraw $3, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1817 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1818 "paddw (%%edi,%%eax,), %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1819 "paddw %%mm2, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1820 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1821 "paddw (%%edi,%%eax,2), %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1822 "paddw %%mm2, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1823 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1824 "movq %%mm5, (%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1825 "paddw %%mm2, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1826 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1827 "movq %%mm1, (%%edi,%%eax,) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1828 "psraw $3, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1829 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1830 "movq %%mm7, (%%edi,%%eax,2) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1831 "addl %%edx, %%edi \n\t" //3*ls |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1832 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1833 "movq "MANGLE(temps)"+1*8, %%mm5 \n\t" //t3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1834 "psraw $3, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1835 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1836 "paddw (%%edi,%%eax,2), %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1837 "psubw %%mm4, %%mm5 \n\t" //d3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1838 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1839 "paddw (%%edi,%%edx,), %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1840 "psraw $3, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1841 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1842 "paddw "MANGLE(temps)"+1*8, %%mm4 \n\t" //d4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1843 "paddw %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1844 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1845 "paddw (%%edi,%%eax,4), %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1846 "paddw %%mm2, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1847 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1848 "movq %%mm0, (%%edi,%%eax,2) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1849 "psraw $3, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1850 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1851 "paddw (%%edi), %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1852 "psraw $3, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1853 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1854 "paddw (%%edi,%%eax,), %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1855 "addl $"DCTSIZE_S"*2*4, %%esi \n\t" //4 rows |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1856 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1857 "movq %%mm3, (%%edi,%%edx,) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1858 "movq %%mm6, (%%edi,%%eax,4) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1859 "movq %%mm5, (%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1860 "movq %%mm4, (%%edi,%%eax,) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1861 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1862 "subl %%edx, %%edi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1863 "addl $8, %%edi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1864 "decl %%ecx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1865 "jnz 1b \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1866 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1867 : "+S"(workspace), "+D"(output_adr), "+c"(cnt) //input regs |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1868 : "a"(output_stride*sizeof(short)) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1869 : "%edx" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1870 ); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1871 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1872 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1873 #endif // HAVE_MMX |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1874 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1875 #ifndef HAVE_MMX |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1876 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1877 static void row_fdct_c(DCTELEM *data, const uint8_t *pixels, int line_size, int cnt) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1878 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1879 int_simd16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1880 int_simd16_t tmp10, tmp11, tmp12, tmp13; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1881 int_simd16_t z1, z2, z3, z4, z5, z11, z13; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1882 DCTELEM *dataptr; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1883 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1884 cnt*=4; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1885 // Pass 1: process rows. |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1886 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1887 dataptr = data; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1888 for (; cnt > 0; cnt--) { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1889 tmp0 = pixels[line_size*0] + pixels[line_size*7]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1890 tmp7 = pixels[line_size*0] - pixels[line_size*7]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1891 tmp1 = pixels[line_size*1] + pixels[line_size*6]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1892 tmp6 = pixels[line_size*1] - pixels[line_size*6]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1893 tmp2 = pixels[line_size*2] + pixels[line_size*5]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1894 tmp5 = pixels[line_size*2] - pixels[line_size*5]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1895 tmp3 = pixels[line_size*3] + pixels[line_size*4]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1896 tmp4 = pixels[line_size*3] - pixels[line_size*4]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1897 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1898 // Even part |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1899 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1900 tmp10 = tmp0 + tmp3; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1901 tmp13 = tmp0 - tmp3; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1902 tmp11 = tmp1 + tmp2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1903 tmp12 = tmp1 - tmp2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1904 //Even columns are written first, this leads to different order of columns |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1905 //in column_fidct(), but they are processed independently, so all ok. |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1906 //Later in the row_idct() columns readed at the same order. |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1907 dataptr[2] = tmp10 + tmp11; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1908 dataptr[3] = tmp10 - tmp11; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1909 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1910 z1 = MULTIPLY16H((tmp12 + tmp13)<<2, FIX_0_707106781); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1911 dataptr[0] = tmp13 + z1; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1912 dataptr[1] = tmp13 - z1; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1913 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1914 // Odd part |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1915 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1916 tmp10 = (tmp4 + tmp5) <<2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1917 tmp11 = (tmp5 + tmp6) <<2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1918 tmp12 = (tmp6 + tmp7) <<2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1919 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1920 z5 = MULTIPLY16H(tmp10 - tmp12, FIX_0_382683433); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1921 z2 = MULTIPLY16H(tmp10, FIX_0_541196100) + z5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1922 z4 = MULTIPLY16H(tmp12, FIX_1_306562965) + z5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1923 z3 = MULTIPLY16H(tmp11, FIX_0_707106781); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1924 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1925 z11 = tmp7 + z3; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1926 z13 = tmp7 - z3; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1927 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1928 dataptr[4] = z13 + z2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1929 dataptr[5] = z13 - z2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1930 dataptr[6] = z11 + z4; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1931 dataptr[7] = z11 - z4; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1932 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1933 pixels++; // advance pointer to next column |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1934 dataptr += DCTSIZE; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1935 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1936 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1937 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1938 #else /* HAVE_MMX */ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1939 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1940 static void row_fdct_mmx(DCTELEM *data, const uint8_t *pixels, int line_size, int cnt) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1941 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1942 asm volatile( |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1943 "leal (%%eax,%%eax,2), %%edx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1944 "6: \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1945 "movd (%%esi), %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1946 "pxor %%mm7, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1947 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1948 "movd (%%esi,%%eax,), %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1949 "punpcklbw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1950 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1951 "movd (%%esi,%%eax,2), %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1952 "punpcklbw %%mm7, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1953 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1954 "punpcklbw %%mm7, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1955 "addl %%edx, %%esi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1956 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1957 "movq %%mm0, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1958 // |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1959 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1960 "movd (%%esi,%%eax,4), %%mm3 \n\t" //7 ;prefetch! |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1961 "movq %%mm1, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1962 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1963 "movd (%%esi,%%edx,), %%mm4 \n\t" //6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1964 "punpcklbw %%mm7, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1965 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1966 "psubw %%mm3, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1967 "punpcklbw %%mm7, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1968 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1969 "paddw %%mm3, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1970 "psubw %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1971 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1972 "movd (%%esi,%%eax,2), %%mm3 \n\t" //5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1973 "paddw %%mm4, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1974 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1975 "movq %%mm5, "MANGLE(temps)"+0*8 \n\t" //t7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1976 "punpcklbw %%mm7, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1977 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1978 "movq %%mm6, "MANGLE(temps)"+1*8 \n\t" //t6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1979 "movq %%mm2, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1980 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1981 "movd (%%esi), %%mm5 \n\t" //3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1982 "paddw %%mm3, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1983 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1984 "movd (%%esi,%%eax,), %%mm6 \n\t" //4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1985 "punpcklbw %%mm7, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1986 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1987 "psubw %%mm3, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1988 "punpcklbw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1989 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1990 "movq %%mm5, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1991 "paddw %%mm6, %%mm5 \n\t" //t3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1992 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1993 "psubw %%mm6, %%mm3 \n\t" //t4 ; t0 t1 t2 t4 t5 t3 - - |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1994 "movq %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1995 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1996 "movq %%mm1, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1997 "psubw %%mm5, %%mm0 \n\t" //t13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1998 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1999 "psubw %%mm2, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2000 "paddw %%mm2, %%mm7 \n\t" //t11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2001 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2002 "paddw %%mm0, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2003 "movq %%mm7, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2004 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2005 "psllw $2, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2006 "paddw %%mm5, %%mm6 \n\t" //t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2007 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2008 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2009 "paddw %%mm6, %%mm7 \n\t" //d2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2010 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2011 "psubw %%mm2, %%mm6 \n\t" //d3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2012 "movq %%mm0, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2013 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2014 //transpose 4x4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2015 "movq %%mm7, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2016 "punpcklwd %%mm6, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2017 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2018 "paddw %%mm1, %%mm0 \n\t" //d0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2019 "punpckhwd %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2020 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2021 "psubw %%mm1, %%mm5 \n\t" //d1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2022 "movq %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2023 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2024 "movq "MANGLE(temps)"+1*8, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2025 "punpcklwd %%mm5, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2026 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2027 "punpckhwd %%mm5, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2028 "movq %%mm0, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2029 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2030 "punpckldq %%mm7, %%mm0 \n\t" //0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2031 "paddw %%mm4, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2032 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2033 "punpckhdq %%mm7, %%mm5 \n\t" //1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2034 "movq %%mm6, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2035 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2036 "movq %%mm0, "DCTSIZE_S"*0*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2037 "punpckldq %%mm2, %%mm6 \n\t" //2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2038 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2039 "movq %%mm5, "DCTSIZE_S"*1*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2040 "punpckhdq %%mm2, %%mm7 \n\t" //3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2041 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2042 "movq %%mm6, "DCTSIZE_S"*2*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2043 "paddw %%mm1, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2044 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2045 "movq %%mm7, "DCTSIZE_S"*3*2(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2046 "psllw $2, %%mm3 \n\t" //t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2047 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2048 "movq "MANGLE(temps)"+0*8, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2049 "psllw $2, %%mm4 \n\t" //t11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2050 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2051 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm4 \n\t" //z3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2052 "paddw %%mm2, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2053 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2054 "psllw $2, %%mm1 \n\t" //t12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2055 "movq %%mm3, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2056 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2057 "pmulhw "MANGLE(MM_FIX_0_541196100)", %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2058 "psubw %%mm1, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2059 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2060 "pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t" //z5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2061 "movq %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2062 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2063 "pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2064 "psubw %%mm4, %%mm2 \n\t" //z13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2065 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2066 "paddw %%mm4, %%mm5 \n\t" //z11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2067 "movq %%mm2, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2068 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2069 "paddw %%mm3, %%mm0 \n\t" //z2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2070 "movq %%mm5, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2071 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2072 "paddw %%mm0, %%mm2 \n\t" //d4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2073 "psubw %%mm0, %%mm6 \n\t" //d5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2074 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2075 "movq %%mm2, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2076 "paddw %%mm3, %%mm1 \n\t" //z4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2077 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2078 //transpose 4x4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2079 "punpcklwd %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2080 "paddw %%mm1, %%mm5 \n\t" //d6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2081 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2082 "punpckhwd %%mm6, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2083 "psubw %%mm1, %%mm7 \n\t" //d7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2084 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2085 "movq %%mm5, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2086 "punpcklwd %%mm7, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2087 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2088 "punpckhwd %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2089 "movq %%mm2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2090 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2091 "punpckldq %%mm5, %%mm2 \n\t" //4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2092 "subl %%edx, %%esi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2093 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2094 "punpckhdq %%mm5, %%mm7 \n\t" //5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2095 "movq %%mm4, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2096 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2097 "movq %%mm2, "DCTSIZE_S"*0*2+"DCTSIZE_S"(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2098 "punpckldq %%mm6, %%mm4 \n\t" //6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2099 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2100 "movq %%mm7, "DCTSIZE_S"*1*2+"DCTSIZE_S"(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2101 "punpckhdq %%mm6, %%mm5 \n\t" //7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2102 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2103 "movq %%mm4, "DCTSIZE_S"*2*2+"DCTSIZE_S"(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2104 "addl $4, %%esi \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2105 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2106 "movq %%mm5, "DCTSIZE_S"*3*2+"DCTSIZE_S"(%%edi) \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2107 "addl $"DCTSIZE_S"*2*4, %%edi \n\t" //4 rows |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2108 "decl %%ecx \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2109 "jnz 6b \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2110 |
15632 | 2111 : "+S"(pixels), "+D"(data), "+c"(cnt) //input regs |
2112 : "a"(line_size) | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2113 : "%edx"); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2114 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2115 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2116 #endif // HAVE_MMX |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2117 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2118 #endif //USE_LIBAVCODEC |