Mercurial > mplayer.hg
annotate libmpcodecs/vf_fspp.c @ 26033:9da3b0c393e6
Better handling of behind-the-camera objects.
Every point that is behind the camera is moved to the clipping plane by
orthographic projection. It is obviously incorrect, but this is a very rare
case, and proper clipping of Bezier curves is not that easy.
author | eugeni |
---|---|
date | Fri, 22 Feb 2008 19:31:49 +0000 |
parents | 15ab840747e2 |
children | 2f15c078db92 |
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 |
17367
401b440a6d76
Update licensing information: The FSF changed postal address.
diego
parents:
17225
diff
changeset
|
17 Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA |
15631
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 |
17012 | 39 #include "config.h" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
40 |
17012 | 41 #include "mp_msg.h" |
42 #include "cpudetect.h" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
43 |
24976 | 44 #include "libavutil/intreadwrite.h" |
25901 | 45 #include "libavutil/mem.h" |
17012 | 46 #include "libavcodec/avcodec.h" |
47 #include "libavcodec/dsputil.h" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
48 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
49 #ifdef HAVE_MALLOC_H |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
50 #include <malloc.h> |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
51 #endif |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
52 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
53 #include "img_format.h" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
54 #include "mp_image.h" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
55 #include "vf.h" |
17012 | 56 #include "libvo/fastmemcpy.h" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
57 |
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 #define BLOCKSZ 12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
60 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
61 static const short custom_threshold[64]= |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
62 // values (296) can't be too high |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
63 // -it causes too big quant dependence |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
64 // or maybe overflow(check), which results in some flashing |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
65 { 71, 296, 295, 237, 71, 40, 38, 19, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
66 245, 193, 185, 121, 102, 73, 53, 27, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
67 158, 129, 141, 107, 97, 73, 50, 26, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
68 102, 116, 109, 98, 82, 66, 45, 23, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
69 71, 94, 95, 81, 70, 56, 38, 20, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
70 56, 77, 74, 66, 56, 44, 30, 15, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
71 38, 53, 50, 45, 38, 30, 21, 11, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
72 20, 27, 26, 23, 20, 15, 11, 5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
73 }; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
74 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
75 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
|
76 { 0, 48, 12, 60, 3, 51, 15, 63, }, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
77 { 32, 16, 44, 28, 35, 19, 47, 31, }, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
78 { 8, 56, 4, 52, 11, 59, 7, 55, }, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
79 { 40, 24, 36, 20, 43, 27, 39, 23, }, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
80 { 2, 50, 14, 62, 1, 49, 13, 61, }, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
81 { 34, 18, 46, 30, 33, 17, 45, 29, }, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
82 { 10, 58, 6, 54, 9, 57, 5, 53, }, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
83 { 42, 26, 38, 22, 41, 25, 37, 21, }, |
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 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
86 struct vf_priv_s { //align 16 ! |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
87 uint64_t threshold_mtx_noq[8*2]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
88 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
|
89 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
90 int log2_count; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
91 int temp_stride; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
92 int qp; |
17225 | 93 int mpeg2; |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
94 int prev_q; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
95 uint8_t *src; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
96 int16_t *temp; |
17225 | 97 int bframes; |
17133
a2b24e0d7772
prevent flicker on b-frames, trivial port from vf_spp
henry
parents:
17012
diff
changeset
|
98 char *non_b_qp; |
15631
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 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
101 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
102 #ifndef HAVE_MMX |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
103 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
104 //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
|
105 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
|
106 {int y, x; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
107 #define STORE(pos) \ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
108 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
|
109 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
|
110 if(temp & 0x100) temp= ~(temp>>31); \ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
111 dst[x + pos]= temp; |
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 for(y=0; y<height; y++){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
114 const uint8_t *d= dither[y]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
115 for(x=0; x<width; x+=8){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
116 int temp; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
117 STORE(0); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
118 STORE(1); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
119 STORE(2); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
120 STORE(3); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
121 STORE(4); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
122 STORE(5); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
123 STORE(6); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
124 STORE(7); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
125 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
126 src+=src_stride; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
127 dst+=dst_stride; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
128 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
129 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
130 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
131 //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
|
132 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
|
133 {int y, x; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
134 #define STORE2(pos) \ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
135 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
|
136 src[x + pos + 16*src_stride]=0; \ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
137 if(temp & 0x100) temp= ~(temp>>31); \ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
138 dst[x + pos]= temp; |
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 for(y=0; y<height; y++){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
141 const uint8_t *d= dither[y]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
142 for(x=0; x<width; x+=8){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
143 int temp; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
144 STORE2(0); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
145 STORE2(1); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
146 STORE2(2); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
147 STORE2(3); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
148 STORE2(4); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
149 STORE2(5); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
150 STORE2(6); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
151 STORE2(7); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
152 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
153 src+=src_stride; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
154 dst+=dst_stride; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
155 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
156 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
157 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
158 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
|
159 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
160 int a; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
161 for(a=0;a<64;a++) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
162 ((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
|
163 } |
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 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
|
166 static void row_idct_c(DCTELEM* workspace, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
167 int16_t* output_adr, int output_stride, int cnt); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
168 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
|
169 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
170 //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
|
171 #define store_slice_s store_slice_c |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
172 #define store_slice2_s store_slice2_c |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
173 #define mul_thrmat_s mul_thrmat_c |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
174 #define column_fidct_s column_fidct_c |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
175 #define row_idct_s row_idct_c |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
176 #define row_fdct_s row_fdct_c |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
177 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
178 #else /* HAVE_MMX */ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
179 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
180 //This func reads from 1 slice, 1 and clears 0 & 1 |
15634 | 181 static void store_slice_mmx(uint8_t *dst, int16_t *src, long dst_stride, long src_stride, long width, long height, long log2_scale) |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
182 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
183 const uint8_t *od=&dither[0][0]; |
15634 | 184 const uint8_t *end=&dither[height][0]; |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
185 width = (width+7)&~7; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
186 dst_stride-=width; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
187 //src_stride=(src_stride-width)*2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
188 asm volatile( |
15634 | 189 "mov %5, %%"REG_d" \n\t" |
190 "mov %6, %%"REG_S" \n\t" | |
191 "mov %7, %%"REG_D" \n\t" | |
192 "mov %1, %%"REG_a" \n\t" | |
193 "movd %%"REG_d", %%mm5 \n\t" | |
194 "xor $-1, %%"REG_d" \n\t" | |
195 "mov %%"REG_a", %%"REG_c" \n\t" | |
196 "add $7, %%"REG_d" \n\t" | |
197 "neg %%"REG_a" \n\t" | |
198 "sub %0, %%"REG_c" \n\t" | |
199 "add %%"REG_c", %%"REG_c" \n\t" | |
200 "movd %%"REG_d", %%mm2 \n\t" | |
201 "mov %%"REG_c", %1 \n\t" | |
202 "mov %2, %%"REG_d" \n\t" | |
203 "shl $4, %%"REG_a" \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
204 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
205 "2: \n\t" |
15634 | 206 "movq (%%"REG_d"), %%mm3 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
207 "movq %%mm3, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
208 "pxor %%mm7, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
209 "punpcklbw %%mm7, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
210 "punpckhbw %%mm7, %%mm4 \n\t" |
15634 | 211 "mov %0, %%"REG_c" \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
212 "psraw %%mm5, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
213 "psraw %%mm5, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
214 "1: \n\t" |
15634 | 215 "movq %%mm7, (%%"REG_S",%%"REG_a",) \n\t" |
216 "movq (%%"REG_S"), %%mm0 \n\t" | |
217 "movq 8(%%"REG_S"), %%mm1 \n\t" | |
218 | |
219 "movq %%mm7, 8(%%"REG_S",%%"REG_a",) \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
220 "paddw %%mm3, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
221 "paddw %%mm4, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
222 |
15634 | 223 "movq %%mm7, (%%"REG_S") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
224 "psraw %%mm2, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
225 "psraw %%mm2, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
226 |
15634 | 227 "movq %%mm7, 8(%%"REG_S") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
228 "packuswb %%mm1, %%mm0 \n\t" |
15634 | 229 "add $16, %%"REG_S" \n\t" |
230 | |
231 "movq %%mm0, (%%"REG_D") \n\t" | |
232 "add $8, %%"REG_D" \n\t" | |
233 "sub $8, %%"REG_c" \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
234 "jg 1b \n\t" |
15634 | 235 "add %1, %%"REG_S" \n\t" |
236 "add $8, %%"REG_d" \n\t" | |
237 "add %3, %%"REG_D" \n\t" | |
238 "cmp %4, %%"REG_d" \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
239 "jl 2b \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
240 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
241 : |
15634 | 242 : "m" (width), "m" (src_stride), "g" (od), "m" (dst_stride), "g" (end), |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
243 "m" (log2_scale), "m" (src), "m" (dst) //input |
15634 | 244 : "%"REG_a, "%"REG_c, "%"REG_d, "%"REG_S, "%"REG_D |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
245 ); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
246 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
247 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
248 //This func reads from 2 slices, 0 & 2 and clears 2-nd |
15634 | 249 static void store_slice2_mmx(uint8_t *dst, int16_t *src, long dst_stride, long src_stride, long width, long height, long log2_scale) |
15631
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 const uint8_t *od=&dither[0][0]; |
15634 | 252 const uint8_t *end=&dither[height][0]; |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
253 width = (width+7)&~7; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
254 dst_stride-=width; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
255 //src_stride=(src_stride-width)*2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
256 asm volatile( |
15634 | 257 "mov %5, %%"REG_d" \n\t" |
258 "mov %6, %%"REG_S" \n\t" | |
259 "mov %7, %%"REG_D" \n\t" | |
260 "mov %1, %%"REG_a" \n\t" | |
261 "movd %%"REG_d", %%mm5 \n\t" | |
262 "xor $-1, %%"REG_d" \n\t" | |
263 "mov %%"REG_a", %%"REG_c" \n\t" | |
264 "add $7, %%"REG_d" \n\t" | |
265 "sub %0, %%"REG_c" \n\t" | |
266 "add %%"REG_c", %%"REG_c" \n\t" | |
267 "movd %%"REG_d", %%mm2 \n\t" | |
268 "mov %%"REG_c", %1 \n\t" | |
269 "mov %2, %%"REG_d" \n\t" | |
270 "shl $5, %%"REG_a" \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
271 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
272 "2: \n\t" |
15634 | 273 "movq (%%"REG_d"), %%mm3 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
274 "movq %%mm3, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
275 "pxor %%mm7, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
276 "punpcklbw %%mm7, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
277 "punpckhbw %%mm7, %%mm4 \n\t" |
15634 | 278 "mov %0, %%"REG_c" \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
279 "psraw %%mm5, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
280 "psraw %%mm5, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
281 "1: \n\t" |
15634 | 282 "movq (%%"REG_S"), %%mm0 \n\t" |
283 "movq 8(%%"REG_S"), %%mm1 \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
284 "paddw %%mm3, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
285 |
15634 | 286 "paddw (%%"REG_S",%%"REG_a",), %%mm0 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
287 "paddw %%mm4, %%mm1 \n\t" |
15634 | 288 "movq 8(%%"REG_S",%%"REG_a",), %%mm6 \n\t" |
289 | |
290 "movq %%mm7, (%%"REG_S",%%"REG_a",) \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
291 "psraw %%mm2, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
292 "paddw %%mm6, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
293 |
15634 | 294 "movq %%mm7, 8(%%"REG_S",%%"REG_a",) \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
295 "psraw %%mm2, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
296 "packuswb %%mm1, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
297 |
15634 | 298 "movq %%mm0, (%%"REG_D") \n\t" |
299 "add $16, %%"REG_S" \n\t" | |
300 "add $8, %%"REG_D" \n\t" | |
301 "sub $8, %%"REG_c" \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
302 "jg 1b \n\t" |
15634 | 303 "add %1, %%"REG_S" \n\t" |
304 "add $8, %%"REG_d" \n\t" | |
305 "add %3, %%"REG_D" \n\t" | |
306 "cmp %4, %%"REG_d" \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
307 "jl 2b \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
308 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
309 : |
15634 | 310 : "m" (width), "m" (src_stride), "g" (od), "m" (dst_stride), "g" (end), |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
311 "m" (log2_scale), "m" (src), "m" (dst) //input |
15634 | 312 : "%"REG_a, "%"REG_c, "%"REG_d, "%"REG_D, "%"REG_S |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
313 ); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
314 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
315 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
316 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
|
317 { |
15634 | 318 uint64_t *adr=&p->threshold_mtx_noq[0]; |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
319 asm volatile( |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
320 "movd %0, %%mm7 \n\t" |
15634 | 321 "add $8*8*2, %%"REG_D" \n\t" |
322 "movq 0*8(%%"REG_S"), %%mm0 \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
323 "punpcklwd %%mm7, %%mm7 \n\t" |
15634 | 324 "movq 1*8(%%"REG_S"), %%mm1 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
325 "punpckldq %%mm7, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
326 "pmullw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
327 |
15634 | 328 "movq 2*8(%%"REG_S"), %%mm2 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
329 "pmullw %%mm7, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
330 |
15634 | 331 "movq 3*8(%%"REG_S"), %%mm3 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
332 "pmullw %%mm7, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
333 |
15634 | 334 "movq %%mm0, 0*8(%%"REG_D") \n\t" |
335 "movq 4*8(%%"REG_S"), %%mm4 \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
336 "pmullw %%mm7, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
337 |
15634 | 338 "movq %%mm1, 1*8(%%"REG_D") \n\t" |
339 "movq 5*8(%%"REG_S"), %%mm5 \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
340 "pmullw %%mm7, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
341 |
15634 | 342 "movq %%mm2, 2*8(%%"REG_D") \n\t" |
343 "movq 6*8(%%"REG_S"), %%mm6 \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
344 "pmullw %%mm7, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
345 |
15634 | 346 "movq %%mm3, 3*8(%%"REG_D") \n\t" |
347 "movq 7*8+0*8(%%"REG_S"), %%mm0 \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
348 "pmullw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
349 |
15634 | 350 "movq %%mm4, 4*8(%%"REG_D") \n\t" |
351 "movq 7*8+1*8(%%"REG_S"), %%mm1 \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
352 "pmullw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
353 |
15634 | 354 "movq %%mm5, 5*8(%%"REG_D") \n\t" |
355 "movq 7*8+2*8(%%"REG_S"), %%mm2 \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
356 "pmullw %%mm7, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
357 |
15634 | 358 "movq %%mm6, 6*8(%%"REG_D") \n\t" |
359 "movq 7*8+3*8(%%"REG_S"), %%mm3 \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
360 "pmullw %%mm7, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
361 |
15634 | 362 "movq %%mm0, 7*8+0*8(%%"REG_D") \n\t" |
363 "movq 7*8+4*8(%%"REG_S"), %%mm4 \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
364 "pmullw %%mm7, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
365 |
15634 | 366 "movq %%mm1, 7*8+1*8(%%"REG_D") \n\t" |
367 "movq 7*8+5*8(%%"REG_S"), %%mm5 \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
368 "pmullw %%mm7, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
369 |
15634 | 370 "movq %%mm2, 7*8+2*8(%%"REG_D") \n\t" |
371 "movq 7*8+6*8(%%"REG_S"), %%mm6 \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
372 "pmullw %%mm7, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
373 |
15634 | 374 "movq %%mm3, 7*8+3*8(%%"REG_D") \n\t" |
375 "movq 14*8+0*8(%%"REG_S"), %%mm0 \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
376 "pmullw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
377 |
15634 | 378 "movq %%mm4, 7*8+4*8(%%"REG_D") \n\t" |
379 "movq 14*8+1*8(%%"REG_S"), %%mm1 \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
380 "pmullw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
381 |
15634 | 382 "movq %%mm5, 7*8+5*8(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
383 "pmullw %%mm7, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
384 |
15634 | 385 "movq %%mm6, 7*8+6*8(%%"REG_D") \n\t" |
386 "movq %%mm0, 14*8+0*8(%%"REG_D") \n\t" | |
387 "movq %%mm1, 14*8+1*8(%%"REG_D") \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
388 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
389 : "+g" (q), "+S" (adr), "+D" (adr) |
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 ); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
392 } |
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 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
|
395 static void row_idct_mmx(DCTELEM* workspace, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
396 int16_t* output_adr, int output_stride, int cnt); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
397 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
|
398 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
399 #define store_slice_s store_slice_mmx |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
400 #define store_slice2_s store_slice2_mmx |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
401 #define mul_thrmat_s mul_thrmat_mmx |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
402 #define column_fidct_s column_fidct_mmx |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
403 #define row_idct_s row_idct_mmx |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
404 #define row_fdct_s row_fdct_mmx |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
405 #endif // HAVE_MMX |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
406 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
407 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
|
408 int dst_stride, int src_stride, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
409 int width, int height, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
410 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
|
411 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
412 int x, x0, y, es, qy, t; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
413 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
|
414 const int step=6-p->log2_count; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
415 const int qps= 3 + is_luma; |
15634 | 416 int32_t __attribute__((aligned(32))) block_align[4*8*BLOCKSZ+ 4*8*BLOCKSZ]; |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
417 DCTELEM *block= (DCTELEM *)block_align; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
418 DCTELEM *block3=(DCTELEM *)(block_align+4*8*BLOCKSZ); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
419 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
420 memset(block3, 0, 4*8*BLOCKSZ); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
421 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
422 //p->src=src-src_stride*8-8;//! |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
423 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
|
424 for(y=0; y<height; y++){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
425 int index= 8 + 8*stride + y*stride; |
23457
a124f3abc1ec
Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents:
21578
diff
changeset
|
426 fast_memcpy(p->src + index, src + y*src_stride, width);//this line can be avoided by using DR & user fr.buffers |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
427 for(x=0; x<8; x++){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
428 p->src[index - x - 1]= p->src[index + x ]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
429 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
|
430 } |
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 for(y=0; y<8; y++){ |
23457
a124f3abc1ec
Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents:
21578
diff
changeset
|
433 fast_memcpy(p->src + ( 7-y)*stride, p->src + ( y+8)*stride, stride); |
a124f3abc1ec
Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents:
21578
diff
changeset
|
434 fast_memcpy(p->src + (height+8+y)*stride, p->src + (height-y+7)*stride, stride); |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
435 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
436 //FIXME (try edge emu) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
437 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
438 for(y=8; y<24; y++) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
439 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
|
440 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
441 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
|
442 qy=y-4; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
443 if (qy>height-1) qy=height-1; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
444 if (qy<0) qy=0; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
445 qy=(qy>>qps)*qp_stride; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
446 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
|
447 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
|
448 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
|
449 if(p->qp) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
450 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
|
451 else |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
452 for (x=0; x<8*(BLOCKSZ-1); x+=8) { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
453 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
|
454 if (t<0) t=0;//t always < width-2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
455 t=qp_store[qy+(t>>qps)]; |
17225 | 456 if(p->mpeg2) t>>=1; //copy p->mpeg2,prev_q to locals? |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
457 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
|
458 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
|
459 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
460 row_idct_s(block3+0*8, p->temp + (y&15)*stride+x0+2-(y&1), stride, 2*(BLOCKSZ-1)); |
25568 | 461 memmove(block, block+(BLOCKSZ-1)*64, 8*8*sizeof(DCTELEM)); //cycling |
462 memmove(block3, block3+(BLOCKSZ-1)*64, 6*8*sizeof(DCTELEM)); | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
463 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
464 // |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
465 es=width+8-x0; // 8, ... |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
466 if (es>8) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
467 row_fdct_s(block+8*8, p->src + y*stride+8+x0 +2-(y&1), stride, (es-4)>>2); |
21578 | 468 column_fidct_s((int16_t*)(&p->threshold_mtx[0]), block, block3, es&(~1)); |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
469 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
|
470 {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
|
471 if (!(y1&7) && y1) { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
472 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
|
473 dst_stride, stride, width, 8, 5-p->log2_count); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
474 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
|
475 dst_stride, stride, width, 8, 5-p->log2_count); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
476 } } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
477 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
478 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
479 if (y&7) { // == height & 7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
480 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
|
481 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
|
482 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
|
483 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
|
484 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
485 } |
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 static int config(struct vf_instance_s* vf, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
488 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
|
489 unsigned int flags, unsigned int outfmt) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
490 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
491 int h= (height+16+15)&(~15); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
492 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
493 vf->priv->temp_stride= (width+16+15)&(~15); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
494 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
|
495 //this can also be avoided, see above |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
496 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
|
497 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
498 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
|
499 } |
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 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
|
502 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
503 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
|
504 // 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
|
505 vf->dmpi=vf_get_image(vf->next,mpi->imgfmt, |
16018
bdf1b4ecb906
use stored dimensions instead of visible one when (vf_)get_image is called
iive
parents:
15651
diff
changeset
|
506 mpi->type, mpi->flags, mpi->width, mpi->height); |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
507 mpi->planes[0]=vf->dmpi->planes[0]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
508 mpi->stride[0]=vf->dmpi->stride[0]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
509 mpi->width=vf->dmpi->width; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
510 if(mpi->flags&MP_IMGFLAG_PLANAR){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
511 mpi->planes[1]=vf->dmpi->planes[1]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
512 mpi->planes[2]=vf->dmpi->planes[2]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
513 mpi->stride[1]=vf->dmpi->stride[1]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
514 mpi->stride[2]=vf->dmpi->stride[2]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
515 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
516 mpi->flags|=MP_IMGFLAG_DIRECT; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
517 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
518 |
17906
20aca9baf5d8
passing pts through the filter layer (lets see if pts or cola comes out at the end)
michael
parents:
17523
diff
changeset
|
519 static int put_image(struct vf_instance_s* vf, mp_image_t *mpi, double pts) |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
520 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
521 mp_image_t *dmpi; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
522 if(!(mpi->flags&MP_IMGFLAG_DIRECT)){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
523 // 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
|
524 dmpi=vf_get_image(vf->next,mpi->imgfmt, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
525 MP_IMGTYPE_TEMP, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
526 MP_IMGFLAG_ACCEPT_STRIDE|MP_IMGFLAG_PREFER_ALIGNED_STRIDE, |
16018
bdf1b4ecb906
use stored dimensions instead of visible one when (vf_)get_image is called
iive
parents:
15651
diff
changeset
|
527 mpi->width,mpi->height); |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
528 vf_clone_mpi_attributes(dmpi, mpi); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
529 }else{ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
530 dmpi=vf->dmpi; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
531 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
532 |
17225 | 533 vf->priv->mpeg2= mpi->qscale_type; |
17133
a2b24e0d7772
prevent flicker on b-frames, trivial port from vf_spp
henry
parents:
17012
diff
changeset
|
534 if(mpi->pict_type != 3 && mpi->qscale && !vf->priv->qp){ |
a2b24e0d7772
prevent flicker on b-frames, trivial port from vf_spp
henry
parents:
17012
diff
changeset
|
535 if(!vf->priv->non_b_qp) |
20585 | 536 vf->priv->non_b_qp= malloc(mpi->qstride * ((mpi->h + 15) >> 4)); |
23457
a124f3abc1ec
Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents:
21578
diff
changeset
|
537 fast_memcpy(vf->priv->non_b_qp, mpi->qscale, mpi->qstride * ((mpi->h + 15) >> 4)); |
17133
a2b24e0d7772
prevent flicker on b-frames, trivial port from vf_spp
henry
parents:
17012
diff
changeset
|
538 } |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
539 if(vf->priv->log2_count || !(mpi->flags&MP_IMGFLAG_DIRECT)){ |
17133
a2b24e0d7772
prevent flicker on b-frames, trivial port from vf_spp
henry
parents:
17012
diff
changeset
|
540 char *qp_tab= vf->priv->non_b_qp; |
17225 | 541 if(vf->priv->bframes || !qp_tab) |
17133
a2b24e0d7772
prevent flicker on b-frames, trivial port from vf_spp
henry
parents:
17012
diff
changeset
|
542 qp_tab= mpi->qscale; |
a2b24e0d7772
prevent flicker on b-frames, trivial port from vf_spp
henry
parents:
17012
diff
changeset
|
543 |
17225 | 544 if(qp_tab || vf->priv->qp){ |
17133
a2b24e0d7772
prevent flicker on b-frames, trivial port from vf_spp
henry
parents:
17012
diff
changeset
|
545 filter(vf->priv, dmpi->planes[0], mpi->planes[0], dmpi->stride[0], mpi->stride[0], |
a2b24e0d7772
prevent flicker on b-frames, trivial port from vf_spp
henry
parents:
17012
diff
changeset
|
546 mpi->w, mpi->h, qp_tab, mpi->qstride, 1); |
a2b24e0d7772
prevent flicker on b-frames, trivial port from vf_spp
henry
parents:
17012
diff
changeset
|
547 filter(vf->priv, dmpi->planes[1], mpi->planes[1], dmpi->stride[1], mpi->stride[1], |
a2b24e0d7772
prevent flicker on b-frames, trivial port from vf_spp
henry
parents:
17012
diff
changeset
|
548 mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, qp_tab, mpi->qstride, 0); |
a2b24e0d7772
prevent flicker on b-frames, trivial port from vf_spp
henry
parents:
17012
diff
changeset
|
549 filter(vf->priv, dmpi->planes[2], mpi->planes[2], dmpi->stride[2], mpi->stride[2], |
a2b24e0d7772
prevent flicker on b-frames, trivial port from vf_spp
henry
parents:
17012
diff
changeset
|
550 mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, qp_tab, mpi->qstride, 0); |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
551 }else{ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
552 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
|
553 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
|
554 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
|
555 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
556 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
557 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
558 #ifdef HAVE_MMX |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
559 if(gCpuCaps.hasMMX) asm volatile ("emms\n\t"); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
560 #endif |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
561 #ifdef HAVE_MMX2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
562 if(gCpuCaps.hasMMX2) asm volatile ("sfence\n\t"); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
563 #endif |
17906
20aca9baf5d8
passing pts through the filter layer (lets see if pts or cola comes out at the end)
michael
parents:
17523
diff
changeset
|
564 return vf_next_put_image(vf,dmpi, pts); |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
565 } |
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 static void uninit(struct vf_instance_s* vf) |
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) return; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
570 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
571 if(vf->priv->temp) av_free(vf->priv->temp); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
572 vf->priv->temp= NULL; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
573 if(vf->priv->src) av_free(vf->priv->src); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
574 vf->priv->src= NULL; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
575 //if(vf->priv->avctx) free(vf->priv->avctx); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
576 //vf->priv->avctx= NULL; |
17133
a2b24e0d7772
prevent flicker on b-frames, trivial port from vf_spp
henry
parents:
17012
diff
changeset
|
577 if(vf->priv->non_b_qp) free(vf->priv->non_b_qp); |
a2b24e0d7772
prevent flicker on b-frames, trivial port from vf_spp
henry
parents:
17012
diff
changeset
|
578 vf->priv->non_b_qp= NULL; |
15631
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 av_free(vf->priv); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
581 vf->priv=NULL; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
582 } |
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 //===========================================================================// |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
585 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
586 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
|
587 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
588 switch(fmt){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
589 case IMGFMT_YVU9: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
590 case IMGFMT_IF09: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
591 case IMGFMT_YV12: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
592 case IMGFMT_I420: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
593 case IMGFMT_IYUV: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
594 case IMGFMT_CLPL: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
595 case IMGFMT_Y800: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
596 case IMGFMT_Y8: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
597 case IMGFMT_444P: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
598 case IMGFMT_422P: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
599 case IMGFMT_411P: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
600 return vf_next_query_format(vf,fmt); |
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 return 0; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
603 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
604 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
605 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
|
606 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
607 switch(request){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
608 case VFCTRL_QUERY_MAX_PP_LEVEL: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
609 return 5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
610 case VFCTRL_SET_PP_LEVEL: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
611 vf->priv->log2_count= *((unsigned int*)data); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
612 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
|
613 return CONTROL_TRUE; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
614 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
615 return vf_next_control(vf,request,data); |
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 open(vf_instance_t *vf, char* args) |
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 int i=0, bias; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
621 int custom_threshold_m[64]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
622 int log2c=-1; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
623 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
624 vf->config=config; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
625 vf->put_image=put_image; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
626 vf->get_image=get_image; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
627 vf->query_format=query_format; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
628 vf->uninit=uninit; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
629 vf->control= control; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
630 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
|
631 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
632 avcodec_init(); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
633 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
634 //vf->priv->avctx= avcodec_alloc_context(); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
635 //dsputil_init(&vf->priv->dsp, vf->priv->avctx); |
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->priv->log2_count= 4; |
17225 | 638 vf->priv->bframes = 0; |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
639 |
17225 | 640 if (args) sscanf(args, "%d:%d:%d:%d", &log2c, &vf->priv->qp, &i, &vf->priv->bframes); |
641 | |
642 if( log2c >=4 && log2c <=5 ) | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
643 vf->priv->log2_count = log2c; |
15651
6a0494e09435
sanity checks for options; treat quality > 5 as 5, not 4
henry
parents:
15634
diff
changeset
|
644 else if( log2c >= 6 ) |
6a0494e09435
sanity checks for options; treat quality > 5 as 5, not 4
henry
parents:
15634
diff
changeset
|
645 vf->priv->log2_count = 5; |
15631
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 if(vf->priv->qp < 0) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
648 vf->priv->qp = 0; |
15651
6a0494e09435
sanity checks for options; treat quality > 5 as 5, not 4
henry
parents:
15634
diff
changeset
|
649 |
6a0494e09435
sanity checks for options; treat quality > 5 as 5, not 4
henry
parents:
15634
diff
changeset
|
650 if (i < -15) i = -15; |
6a0494e09435
sanity checks for options; treat quality > 5 as 5, not 4
henry
parents:
15634
diff
changeset
|
651 if (i > 32) i = 32; |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
652 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
653 bias= (1<<4)+i; //regulable |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
654 vf->priv->prev_q=0; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
655 // |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
656 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
|
657 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
|
658 for(i=0;i<8;i++){ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
659 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
|
660 |(((uint64_t)custom_threshold_m[i*8+6])<<16) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
661 |(((uint64_t)custom_threshold_m[i*8+0])<<32) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
662 |(((uint64_t)custom_threshold_m[i*8+4])<<48); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
663 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
|
664 |(((uint64_t)custom_threshold_m[i*8+3])<<16) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
665 |(((uint64_t)custom_threshold_m[i*8+1])<<32) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
666 |(((uint64_t)custom_threshold_m[i*8+7])<<48); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
667 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
668 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
669 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
|
670 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
671 return 1; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
672 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
673 |
25221 | 674 const vf_info_t vf_info_fspp = { |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
675 "fast simple postprocess", |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
676 "fspp", |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
677 "Michael Niedermayer, Nikolaj Poroshin", |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
678 "", |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
679 open, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
680 NULL |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
681 }; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
682 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
683 //==================================================================== |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
684 //Specific spp's dct, idct and threshold functions |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
685 //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
|
686 |
17012 | 687 #include "mangle.h" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
688 //#define MANGLE(a) #a |
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 //typedef int16_t DCTELEM; //! only int16_t |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
691 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
692 #define DCTSIZE 8 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
693 #define DCTSIZE_S "8" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
694 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
695 #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
|
696 #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
|
697 #define FIX64(x,s) C64(FIX(x,s)) |
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 MULTIPLY16H(x,k) (((x)*(k))>>16) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
700 #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
|
701 #define DESCALE(x,n) (((x) + (1 << ((n)-1))) >> n) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
702 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
703 #ifdef HAVE_MMX |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
704 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
705 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
|
706 |
25901 | 707 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_382683433)=FIX64(0.382683433, 14); |
708 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_541196100)=FIX64(0.541196100, 14); | |
709 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_707106781)=FIX64(0.707106781, 14); | |
710 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_306562965)=FIX64(1.306562965, 14); | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
711 |
25901 | 712 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_414213562_A)=FIX64(1.414213562, 14); |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
713 |
25901 | 714 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_847759065)=FIX64(1.847759065, 13); |
715 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_2_613125930)=FIX64(-2.613125930, 13); //- | |
716 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_414213562)=FIX64(1.414213562, 13); | |
717 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_082392200)=FIX64(1.082392200, 13); | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
718 //for t3,t5,t7 == 0 shortcut |
25901 | 719 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_847759065)=FIX64(0.847759065, 14); |
720 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_566454497)=FIX64(0.566454497, 14); | |
721 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_198912367)=FIX64(0.198912367, 14); | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
722 |
25901 | 723 DECLARE_ASM_CONST(8, uint64_t, MM_DESCALE_RND)=C64(4); |
724 DECLARE_ASM_CONST(8, uint64_t, MM_2)=C64(2); | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
725 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
726 #else /* !HAVE_MMX */ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
727 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
728 typedef int32_t int_simd16_t; |
25902 | 729 static const int16_t FIX_0_382683433=FIX(0.382683433, 14); |
730 static const int16_t FIX_0_541196100=FIX(0.541196100, 14); | |
731 static const int16_t FIX_0_707106781=FIX(0.707106781, 14); | |
732 static const int16_t FIX_1_306562965=FIX(1.306562965, 14); | |
733 static const int16_t FIX_1_414213562_A=FIX(1.414213562, 14); | |
734 static const int16_t FIX_1_847759065=FIX(1.847759065, 13); | |
735 static const int16_t FIX_2_613125930=FIX(-2.613125930, 13); //- | |
736 static const int16_t FIX_1_414213562=FIX(1.414213562, 13); | |
737 static const int16_t FIX_1_082392200=FIX(1.082392200, 13); | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
738 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
739 #endif |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
740 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
741 #ifndef HAVE_MMX |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
742 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
743 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
|
744 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
745 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
|
746 int_simd16_t tmp10, tmp11, tmp12, tmp13; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
747 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
|
748 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
|
749 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
750 DCTELEM* dataptr; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
751 DCTELEM* wsptr; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
752 int16_t *threshold; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
753 int ctr; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
754 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
755 dataptr = data; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
756 wsptr = output; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
757 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
758 for (; cnt > 0; cnt-=2) { //start positions |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
759 threshold=(int16_t*)thr_adr;//threshold_mtx |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
760 for (ctr = DCTSIZE; ctr > 0; ctr--) { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
761 // Process columns from input, add to output. |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
762 tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
763 tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7]; |
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 tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
766 tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
767 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
768 tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
769 tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
770 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
771 tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
772 tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
773 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
774 // Even part of FDCT |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
775 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
776 tmp10 = tmp0 + tmp3; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
777 tmp13 = tmp0 - tmp3; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
778 tmp11 = tmp1 + tmp2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
779 tmp12 = tmp1 - tmp2; |
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 d0 = tmp10 + tmp11; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
782 d4 = tmp10 - tmp11; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
783 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
784 z1 = MULTIPLY16H((tmp12 + tmp13) <<2, FIX_0_707106781); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
785 d2 = tmp13 + z1; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
786 d6 = tmp13 - z1; |
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 // Even part of IDCT |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
789 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
790 THRESHOLD(tmp0, d0, threshold[0*8]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
791 THRESHOLD(tmp1, d2, threshold[2*8]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
792 THRESHOLD(tmp2, d4, threshold[4*8]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
793 THRESHOLD(tmp3, d6, threshold[6*8]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
794 tmp0+=2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
795 tmp10 = (tmp0 + tmp2)>>2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
796 tmp11 = (tmp0 - tmp2)>>2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
797 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
798 tmp13 = (tmp1 + tmp3)>>2; //+2 ! (psnr decides) |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
799 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
|
800 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
801 tmp0 = tmp10 + tmp13; //->temps |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
802 tmp3 = tmp10 - tmp13; //->temps |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
803 tmp1 = tmp11 + tmp12; //->temps |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
804 tmp2 = tmp11 - tmp12; //->temps |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
805 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
806 // Odd part of FDCT |
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 tmp10 = tmp4 + tmp5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
809 tmp11 = tmp5 + tmp6; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
810 tmp12 = tmp6 + tmp7; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
811 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
812 z5 = MULTIPLY16H((tmp10 - tmp12)<<2, FIX_0_382683433); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
813 z2 = MULTIPLY16H(tmp10 <<2, FIX_0_541196100) + z5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
814 z4 = MULTIPLY16H(tmp12 <<2, FIX_1_306562965) + z5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
815 z3 = MULTIPLY16H(tmp11 <<2, FIX_0_707106781); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
816 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
817 z11 = tmp7 + z3; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
818 z13 = tmp7 - z3; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
819 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
820 d5 = z13 + z2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
821 d3 = z13 - z2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
822 d1 = z11 + z4; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
823 d7 = z11 - z4; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
824 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
825 // Odd part of IDCT |
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 THRESHOLD(tmp4, d1, threshold[1*8]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
828 THRESHOLD(tmp5, d3, threshold[3*8]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
829 THRESHOLD(tmp6, d5, threshold[5*8]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
830 THRESHOLD(tmp7, d7, threshold[7*8]); |
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 //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
|
833 z13 = tmp6 + tmp5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
834 z10 = (tmp6 - tmp5)<<1; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
835 z11 = tmp4 + tmp7; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
836 z12 = (tmp4 - tmp7)<<1; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
837 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
838 tmp7 = (z11 + z13)>>2; //+2 ! |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
839 tmp11 = MULTIPLY16H((z11 - z13)<<1, FIX_1_414213562); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
840 z5 = MULTIPLY16H(z10 + z12, FIX_1_847759065); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
841 tmp10 = MULTIPLY16H(z12, FIX_1_082392200) - z5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
842 tmp12 = MULTIPLY16H(z10, FIX_2_613125930) + z5; // - !! |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
843 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
844 tmp6 = tmp12 - tmp7; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
845 tmp5 = tmp11 - tmp6; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
846 tmp4 = tmp10 + tmp5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
847 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
848 wsptr[DCTSIZE*0]+= (tmp0 + tmp7); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
849 wsptr[DCTSIZE*1]+= (tmp1 + tmp6); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
850 wsptr[DCTSIZE*2]+= (tmp2 + tmp5); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
851 wsptr[DCTSIZE*3]+= (tmp3 - tmp4); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
852 wsptr[DCTSIZE*4]+= (tmp3 + tmp4); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
853 wsptr[DCTSIZE*5]+= (tmp2 - tmp5); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
854 wsptr[DCTSIZE*6]= (tmp1 - tmp6); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
855 wsptr[DCTSIZE*7]= (tmp0 - tmp7); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
856 // |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
857 dataptr++; //next column |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
858 wsptr++; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
859 threshold++; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
860 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
861 dataptr+=8; //skip each second start pos |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
862 wsptr +=8; |
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 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
865 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
866 #else /* HAVE_MMX */ |
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 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
|
869 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
870 asm volatile( |
19372
6334c14b38eb
Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents:
18131
diff
changeset
|
871 ASMALIGN(4) |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
872 "1: \n\t" |
15634 | 873 "movq "DCTSIZE_S"*0*2(%%"REG_S"), %%mm1 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
874 // |
15634 | 875 "movq "DCTSIZE_S"*3*2(%%"REG_S"), %%mm7 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
876 "movq %%mm1, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
877 |
15634 | 878 "paddw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm1 \n\t" //t0 |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
879 "movq %%mm7, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
880 |
15634 | 881 "paddw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm7 \n\t" //t3 |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
882 "movq %%mm1, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
883 |
15634 | 884 "movq "DCTSIZE_S"*1*2(%%"REG_S"), %%mm6 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
885 "psubw %%mm7, %%mm1 \n\t" //t13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
886 |
15634 | 887 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
888 "movq %%mm6, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
889 |
15634 | 890 "paddw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm6 \n\t" //t1 |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
891 "paddw %%mm7, %%mm5 \n\t" //t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
892 |
15634 | 893 "paddw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t2 |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
894 "movq %%mm6, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
895 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
896 "paddw %%mm2, %%mm6 \n\t" //t11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
897 "psubw %%mm2, %%mm7 \n\t" //t12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
898 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
899 "movq %%mm5, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
900 "paddw %%mm6, %%mm5 \n\t" //d0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
901 // i0 t13 t12 i3 i1 d0 - d4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
902 "psubw %%mm6, %%mm2 \n\t" //d4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
903 "paddw %%mm1, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
904 |
15634 | 905 "movq 4*16(%%"REG_d"), %%mm6 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
906 "psllw $2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
907 |
15634 | 908 "psubw 0*16(%%"REG_d"), %%mm5 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
909 "psubw %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
910 |
15634 | 911 "paddusw 0*16(%%"REG_d"), %%mm5 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
912 "paddusw %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
913 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
914 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
915 // |
15634 | 916 "paddw 0*16(%%"REG_d"), %%mm5 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
917 "paddw %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
918 |
15634 | 919 "psubusw 0*16(%%"REG_d"), %%mm5 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
920 "psubusw %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
921 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
922 //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
|
923 // 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
|
924 //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
|
925 "paddw "MANGLE(MM_2)", %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
926 "movq %%mm2, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
927 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
928 "paddw %%mm5, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
929 "psubw %%mm6, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
930 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
931 "movq %%mm1, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
932 "paddw %%mm7, %%mm1 \n\t" //d2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
933 |
15634 | 934 "psubw 2*16(%%"REG_d"), %%mm1 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
935 "psubw %%mm7, %%mm6 \n\t" //d6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
936 |
15634 | 937 "movq 6*16(%%"REG_d"), %%mm7 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
938 "psraw $2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
939 |
15634 | 940 "paddusw 2*16(%%"REG_d"), %%mm1 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
941 "psubw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
942 // t7 d2 /t11 t4 t6 - d6 /t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
943 |
15634 | 944 "paddw 2*16(%%"REG_d"), %%mm1 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
945 "paddusw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
946 |
15634 | 947 "psubusw 2*16(%%"REG_d"), %%mm1 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
948 "paddw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
949 |
15634 | 950 "psubw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm3 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
951 "psubusw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
952 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
953 //movq [edi+"DCTSIZE_S"*2*2], mm1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
954 //movq [edi+"DCTSIZE_S"*6*2], mm6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
955 "movq %%mm1, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
956 "psraw $2, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
957 |
15634 | 958 "psubw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm4 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
959 "psubw %%mm6, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
960 |
15634 | 961 "psubw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm0 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
962 "paddw %%mm7, %%mm6 \n\t" //'t13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
963 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
964 "psraw $2, %%mm6 \n\t" //paddw mm6, MM_2 !! --- |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
965 "movq %%mm2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
966 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
967 "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
|
968 "paddw %%mm6, %%mm2 \n\t" //'t0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
969 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
970 "movq %%mm2, "MANGLE(temps)"+0*8 \n\t" //! |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
971 "psubw %%mm6, %%mm7 \n\t" //'t3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
972 |
15634 | 973 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
974 "psubw %%mm6, %%mm1 \n\t" //'t12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
975 |
15634 | 976 "psubw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t5 |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
977 "movq %%mm5, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
978 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
979 "movq %%mm7, "MANGLE(temps)"+3*8 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
980 "paddw %%mm2, %%mm3 \n\t" //t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
981 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
982 "paddw %%mm4, %%mm2 \n\t" //t11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
983 "paddw %%mm0, %%mm4 \n\t" //t12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
984 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
985 "movq %%mm3, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
986 "psubw %%mm4, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
987 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
988 "psllw $2, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
989 "psllw $2, %%mm7 \n\t" //opt for P6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
990 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
991 "pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
992 "psllw $2, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
993 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
994 "pmulhw "MANGLE(MM_FIX_0_541196100)", %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
995 "psllw $2, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
996 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
997 "pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
998 "paddw %%mm1, %%mm5 \n\t" //'t1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
999 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1000 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1001 "psubw %%mm1, %%mm6 \n\t" //'t2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1002 // t7 't12 't11 t4 t6 - 't13 't10 --- |
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 "paddw %%mm3, %%mm7 \n\t" //z2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1005 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1006 "movq %%mm5, "MANGLE(temps)"+1*8 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1007 "paddw %%mm3, %%mm4 \n\t" //z4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1008 |
15634 | 1009 "movq 3*16(%%"REG_d"), %%mm3 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1010 "movq %%mm0, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1011 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1012 "movq %%mm6, "MANGLE(temps)"+2*8 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1013 "psubw %%mm2, %%mm1 \n\t" //z13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1014 |
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 "paddw %%mm2, %%mm0 \n\t" //z11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1017 "movq %%mm1, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1018 |
15634 | 1019 "movq 5*16(%%"REG_d"), %%mm2 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1020 "psubw %%mm7, %%mm1 \n\t" //d3 |
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 "paddw %%mm7, %%mm5 \n\t" //d5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1023 "psubw %%mm3, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1024 |
15634 | 1025 "movq 1*16(%%"REG_d"), %%mm7 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1026 "psubw %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1027 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1028 "movq %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1029 "paddw %%mm4, %%mm0 \n\t" //d1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1030 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1031 "paddusw %%mm3, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1032 "psubw %%mm4, %%mm6 \n\t" //d7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1033 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1034 // d1 d3 - - - d5 d7 - |
15634 | 1035 "movq 7*16(%%"REG_d"), %%mm4 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1036 "psubw %%mm7, %%mm0 \n\t" |
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 "psubw %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1039 "paddusw %%mm2, %%mm5 \n\t" |
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 "paddusw %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1042 "paddw %%mm3, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1043 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1044 "paddw %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1045 "paddw %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1046 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1047 "psubusw %%mm3, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1048 "psubusw %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1049 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1050 "psubusw %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1051 "movq %%mm1, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1052 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1053 "por %%mm5, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1054 "paddusw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1055 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1056 "por %%mm6, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1057 "paddw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1058 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1059 "packssdw %%mm4, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1060 "psubusw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1061 |
15634 | 1062 "movd %%mm4, %%"REG_a" \n\t" |
1063 "or %%"REG_a", %%"REG_a" \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1064 "jnz 2f \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1065 //movq [edi+"DCTSIZE_S"*3*2], mm1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1066 //movq [edi+"DCTSIZE_S"*5*2], mm5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1067 //movq [edi+"DCTSIZE_S"*1*2], mm0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1068 //movq [edi+"DCTSIZE_S"*7*2], mm6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1069 // t4 t5 - - - t6 t7 - |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1070 //--- t4 (mm0) may be <>0; mm1, mm5, mm6 == 0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1071 //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
|
1072 "movq "MANGLE(temps)"+0*8, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1073 "movq %%mm0, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1074 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1075 "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
|
1076 "movq %%mm1, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1077 |
15634 | 1078 "movq "DCTSIZE_S"*0*2(%%"REG_D"), %%mm5 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1079 "movq %%mm2, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1080 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1081 "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
|
1082 "paddw %%mm4, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1083 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1084 "movq "MANGLE(temps)"+1*8, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1085 //paddw mm3, MM_2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1086 "psraw $2, %%mm3 \n\t" //tmp7 |
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_198912367)", %%mm2 \n\t" //-tmp4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1089 "psubw %%mm3, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1090 |
15634 | 1091 "movq "DCTSIZE_S"*1*2(%%"REG_D"), %%mm7 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1092 "paddw %%mm3, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1093 |
15634 | 1094 "movq %%mm4, "DCTSIZE_S"*7*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1095 "paddw %%mm6, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1096 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1097 "movq "MANGLE(temps)"+2*8, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1098 "psubw %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1099 |
15634 | 1100 "movq "DCTSIZE_S"*2*2(%%"REG_D"), %%mm4 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1101 "paddw %%mm0, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1102 |
15634 | 1103 "movq %%mm5, "DCTSIZE_S"*0*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1104 "paddw %%mm3, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1105 |
15634 | 1106 "movq %%mm6, "DCTSIZE_S"*6*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1107 "psubw %%mm1, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1108 |
15634 | 1109 "movq "DCTSIZE_S"*5*2(%%"REG_D"), %%mm5 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1110 "paddw %%mm1, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1111 |
15634 | 1112 "movq "DCTSIZE_S"*3*2(%%"REG_D"), %%mm6 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1113 "paddw %%mm3, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1114 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1115 "movq "MANGLE(temps)"+3*8, %%mm0 \n\t" |
15634 | 1116 "add $8, %%"REG_S" \n\t" |
1117 | |
1118 "movq %%mm7, "DCTSIZE_S"*1*2(%%"REG_D") \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1119 "paddw %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1120 |
15634 | 1121 "movq %%mm4, "DCTSIZE_S"*2*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1122 "psubw %%mm2, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1123 |
15634 | 1124 "movq "DCTSIZE_S"*4*2(%%"REG_D"), %%mm7 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1125 "paddw %%mm2, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1126 |
15634 | 1127 "movq %%mm5, "DCTSIZE_S"*5*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1128 "paddw %%mm0, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1129 |
15634 | 1130 "movq %%mm6, "DCTSIZE_S"*3*2(%%"REG_D") \n\t" |
1131 | |
1132 "movq %%mm7, "DCTSIZE_S"*4*2(%%"REG_D") \n\t" | |
1133 "add $8, %%"REG_D" \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1134 "jmp 4f \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1135 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1136 "2: \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1137 //--- non DC2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1138 //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
|
1139 //psraw mm5, 2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1140 //psraw mm0, 2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1141 //psraw mm6, 2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1142 "movq %%mm5, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1143 "psubw %%mm1, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1144 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1145 "psllw $1, %%mm5 \n\t" //'z10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1146 "paddw %%mm1, %%mm3 \n\t" //'z13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1147 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1148 "movq %%mm0, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1149 "psubw %%mm6, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1150 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1151 "movq %%mm5, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1152 "psllw $1, %%mm0 \n\t" //'z12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1153 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1154 "pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm1 \n\t" //- |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1155 "paddw %%mm0, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1156 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1157 "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
|
1158 "paddw %%mm6, %%mm2 \n\t" //'z11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1159 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1160 "pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1161 "movq %%mm2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1162 |
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 "movq "MANGLE(temps)"+0*8, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1165 "psubw %%mm3, %%mm2 \n\t" |
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 "psllw $1, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1168 "paddw %%mm3, %%mm7 \n\t" //'t7 |
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 "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
|
1171 "movq %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1172 //paddw mm7, MM_2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1173 "psraw $2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1174 |
15634 | 1175 "paddw "DCTSIZE_S"*0*2(%%"REG_D"), %%mm4 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1176 "psubw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1177 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1178 "movq "MANGLE(temps)"+1*8, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1179 "paddw %%mm7, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1180 |
15634 | 1181 "movq %%mm6, "DCTSIZE_S"*7*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1182 "paddw %%mm5, %%mm1 \n\t" //'t12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1183 |
15634 | 1184 "movq %%mm4, "DCTSIZE_S"*0*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1185 "psubw %%mm7, %%mm1 \n\t" //'t6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1186 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1187 "movq "MANGLE(temps)"+2*8, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1188 "psubw %%mm5, %%mm0 \n\t" //'t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1189 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1190 "movq "MANGLE(temps)"+3*8, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1191 "movq %%mm3, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1192 |
15634 | 1193 "paddw "DCTSIZE_S"*1*2(%%"REG_D"), %%mm3 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1194 "psubw %%mm1, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1195 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1196 "psubw %%mm1, %%mm2 \n\t" //'t5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1197 "paddw %%mm1, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1198 |
15634 | 1199 "movq %%mm5, "DCTSIZE_S"*6*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1200 "movq %%mm7, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1201 |
15634 | 1202 "paddw "DCTSIZE_S"*2*2(%%"REG_D"), %%mm7 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1203 "psubw %%mm2, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1204 |
15634 | 1205 "paddw "DCTSIZE_S"*5*2(%%"REG_D"), %%mm4 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1206 "paddw %%mm2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1207 |
15634 | 1208 "movq %%mm3, "DCTSIZE_S"*1*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1209 "paddw %%mm2, %%mm0 \n\t" //'t4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1210 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1211 // 't4 't6 't5 - - - - 't7 |
15634 | 1212 "movq %%mm7, "DCTSIZE_S"*2*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1213 "movq %%mm6, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1214 |
15634 | 1215 "paddw "DCTSIZE_S"*4*2(%%"REG_D"), %%mm6 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1216 "psubw %%mm0, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1217 |
15634 | 1218 "paddw "DCTSIZE_S"*3*2(%%"REG_D"), %%mm1 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1219 "paddw %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1220 |
15634 | 1221 "movq %%mm4, "DCTSIZE_S"*5*2(%%"REG_D") \n\t" |
1222 "add $8, %%"REG_S" \n\t" | |
1223 | |
1224 "movq %%mm6, "DCTSIZE_S"*4*2(%%"REG_D") \n\t" | |
1225 | |
1226 "movq %%mm1, "DCTSIZE_S"*3*2(%%"REG_D") \n\t" | |
1227 "add $8, %%"REG_D" \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1228 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1229 "4: \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1230 //=part 2 (the same)=========================================================== |
15634 | 1231 "movq "DCTSIZE_S"*0*2(%%"REG_S"), %%mm1 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1232 // |
15634 | 1233 "movq "DCTSIZE_S"*3*2(%%"REG_S"), %%mm7 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1234 "movq %%mm1, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1235 |
15634 | 1236 "paddw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm1 \n\t" //t0 |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1237 "movq %%mm7, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1238 |
15634 | 1239 "paddw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm7 \n\t" //t3 |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1240 "movq %%mm1, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1241 |
15634 | 1242 "movq "DCTSIZE_S"*1*2(%%"REG_S"), %%mm6 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1243 "psubw %%mm7, %%mm1 \n\t" //t13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1244 |
15634 | 1245 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1246 "movq %%mm6, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1247 |
15634 | 1248 "paddw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm6 \n\t" //t1 |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1249 "paddw %%mm7, %%mm5 \n\t" //t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1250 |
15634 | 1251 "paddw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t2 |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1252 "movq %%mm6, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1253 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1254 "paddw %%mm2, %%mm6 \n\t" //t11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1255 "psubw %%mm2, %%mm7 \n\t" //t12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1256 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1257 "movq %%mm5, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1258 "paddw %%mm6, %%mm5 \n\t" //d0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1259 // i0 t13 t12 i3 i1 d0 - d4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1260 "psubw %%mm6, %%mm2 \n\t" //d4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1261 "paddw %%mm1, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1262 |
15634 | 1263 "movq 1*8+4*16(%%"REG_d"), %%mm6 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1264 "psllw $2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1265 |
15634 | 1266 "psubw 1*8+0*16(%%"REG_d"), %%mm5 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1267 "psubw %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1268 |
15634 | 1269 "paddusw 1*8+0*16(%%"REG_d"), %%mm5 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1270 "paddusw %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1271 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1272 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1273 // |
15634 | 1274 "paddw 1*8+0*16(%%"REG_d"), %%mm5 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1275 "paddw %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1276 |
15634 | 1277 "psubusw 1*8+0*16(%%"REG_d"), %%mm5 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1278 "psubusw %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1279 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1280 //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
|
1281 // 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
|
1282 //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
|
1283 "paddw "MANGLE(MM_2)", %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1284 "movq %%mm2, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1285 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1286 "paddw %%mm5, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1287 "psubw %%mm6, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1288 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1289 "movq %%mm1, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1290 "paddw %%mm7, %%mm1 \n\t" //d2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1291 |
15634 | 1292 "psubw 1*8+2*16(%%"REG_d"), %%mm1 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1293 "psubw %%mm7, %%mm6 \n\t" //d6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1294 |
15634 | 1295 "movq 1*8+6*16(%%"REG_d"), %%mm7 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1296 "psraw $2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1297 |
15634 | 1298 "paddusw 1*8+2*16(%%"REG_d"), %%mm1 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1299 "psubw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1300 // t7 d2 /t11 t4 t6 - d6 /t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1301 |
15634 | 1302 "paddw 1*8+2*16(%%"REG_d"), %%mm1 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1303 "paddusw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1304 |
15634 | 1305 "psubusw 1*8+2*16(%%"REG_d"), %%mm1 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1306 "paddw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1307 |
15634 | 1308 "psubw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm3 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1309 "psubusw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1310 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1311 //movq [edi+"DCTSIZE_S"*2*2], mm1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1312 //movq [edi+"DCTSIZE_S"*6*2], mm6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1313 "movq %%mm1, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1314 "psraw $2, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1315 |
15634 | 1316 "psubw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm4 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1317 "psubw %%mm6, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1318 |
15634 | 1319 "psubw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm0 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1320 "paddw %%mm7, %%mm6 \n\t" //'t13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1321 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1322 "psraw $2, %%mm6 \n\t" //paddw mm6, MM_2 !! --- |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1323 "movq %%mm2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1324 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1325 "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
|
1326 "paddw %%mm6, %%mm2 \n\t" //'t0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1327 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1328 "movq %%mm2, "MANGLE(temps)"+0*8 \n\t" //! |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1329 "psubw %%mm6, %%mm7 \n\t" //'t3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1330 |
15634 | 1331 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1332 "psubw %%mm6, %%mm1 \n\t" //'t12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1333 |
15634 | 1334 "psubw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t5 |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1335 "movq %%mm5, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1336 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1337 "movq %%mm7, "MANGLE(temps)"+3*8 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1338 "paddw %%mm2, %%mm3 \n\t" //t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1339 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1340 "paddw %%mm4, %%mm2 \n\t" //t11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1341 "paddw %%mm0, %%mm4 \n\t" //t12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1342 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1343 "movq %%mm3, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1344 "psubw %%mm4, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1345 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1346 "psllw $2, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1347 "psllw $2, %%mm7 \n\t" //opt for P6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1348 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1349 "pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1350 "psllw $2, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1351 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1352 "pmulhw "MANGLE(MM_FIX_0_541196100)", %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1353 "psllw $2, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1354 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1355 "pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1356 "paddw %%mm1, %%mm5 \n\t" //'t1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1357 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1358 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1359 "psubw %%mm1, %%mm6 \n\t" //'t2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1360 // t7 't12 't11 t4 t6 - 't13 't10 --- |
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 "paddw %%mm3, %%mm7 \n\t" //z2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1363 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1364 "movq %%mm5, "MANGLE(temps)"+1*8 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1365 "paddw %%mm3, %%mm4 \n\t" //z4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1366 |
15634 | 1367 "movq 1*8+3*16(%%"REG_d"), %%mm3 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1368 "movq %%mm0, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1369 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1370 "movq %%mm6, "MANGLE(temps)"+2*8 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1371 "psubw %%mm2, %%mm1 \n\t" //z13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1372 |
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 "paddw %%mm2, %%mm0 \n\t" //z11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1375 "movq %%mm1, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1376 |
15634 | 1377 "movq 1*8+5*16(%%"REG_d"), %%mm2 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1378 "psubw %%mm7, %%mm1 \n\t" //d3 |
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 "paddw %%mm7, %%mm5 \n\t" //d5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1381 "psubw %%mm3, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1382 |
15634 | 1383 "movq 1*8+1*16(%%"REG_d"), %%mm7 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1384 "psubw %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1385 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1386 "movq %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1387 "paddw %%mm4, %%mm0 \n\t" //d1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1388 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1389 "paddusw %%mm3, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1390 "psubw %%mm4, %%mm6 \n\t" //d7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1391 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1392 // d1 d3 - - - d5 d7 - |
15634 | 1393 "movq 1*8+7*16(%%"REG_d"), %%mm4 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1394 "psubw %%mm7, %%mm0 \n\t" |
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 "psubw %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1397 "paddusw %%mm2, %%mm5 \n\t" |
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 "paddusw %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1400 "paddw %%mm3, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1401 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1402 "paddw %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1403 "paddw %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1404 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1405 "psubusw %%mm3, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1406 "psubusw %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1407 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1408 "psubusw %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1409 "movq %%mm1, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1410 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1411 "por %%mm5, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1412 "paddusw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1413 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1414 "por %%mm6, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1415 "paddw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1416 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1417 "packssdw %%mm4, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1418 "psubusw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1419 |
15634 | 1420 "movd %%mm4, %%"REG_a" \n\t" |
1421 "or %%"REG_a", %%"REG_a" \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1422 "jnz 3f \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1423 //movq [edi+"DCTSIZE_S"*3*2], mm1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1424 //movq [edi+"DCTSIZE_S"*5*2], mm5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1425 //movq [edi+"DCTSIZE_S"*1*2], mm0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1426 //movq [edi+"DCTSIZE_S"*7*2], mm6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1427 // t4 t5 - - - t6 t7 - |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1428 //--- t4 (mm0) may be <>0; mm1, mm5, mm6 == 0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1429 //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
|
1430 "movq "MANGLE(temps)"+0*8, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1431 "movq %%mm0, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1432 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1433 "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
|
1434 "movq %%mm1, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1435 |
15634 | 1436 "movq "DCTSIZE_S"*0*2(%%"REG_D"), %%mm5 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1437 "movq %%mm2, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1438 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1439 "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
|
1440 "paddw %%mm4, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1441 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1442 "movq "MANGLE(temps)"+1*8, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1443 //paddw mm3, MM_2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1444 "psraw $2, %%mm3 \n\t" //tmp7 |
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_198912367)", %%mm2 \n\t" //-tmp4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1447 "psubw %%mm3, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1448 |
15634 | 1449 "movq "DCTSIZE_S"*1*2(%%"REG_D"), %%mm7 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1450 "paddw %%mm3, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1451 |
15634 | 1452 "movq %%mm4, "DCTSIZE_S"*7*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1453 "paddw %%mm6, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1454 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1455 "movq "MANGLE(temps)"+2*8, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1456 "psubw %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1457 |
15634 | 1458 "movq "DCTSIZE_S"*2*2(%%"REG_D"), %%mm4 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1459 "paddw %%mm0, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1460 |
15634 | 1461 "movq %%mm5, "DCTSIZE_S"*0*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1462 "paddw %%mm3, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1463 |
15634 | 1464 "movq %%mm6, "DCTSIZE_S"*6*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1465 "psubw %%mm1, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1466 |
15634 | 1467 "movq "DCTSIZE_S"*5*2(%%"REG_D"), %%mm5 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1468 "paddw %%mm1, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1469 |
15634 | 1470 "movq "DCTSIZE_S"*3*2(%%"REG_D"), %%mm6 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1471 "paddw %%mm3, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1472 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1473 "movq "MANGLE(temps)"+3*8, %%mm0 \n\t" |
15634 | 1474 "add $24, %%"REG_S" \n\t" |
1475 | |
1476 "movq %%mm7, "DCTSIZE_S"*1*2(%%"REG_D") \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1477 "paddw %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1478 |
15634 | 1479 "movq %%mm4, "DCTSIZE_S"*2*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1480 "psubw %%mm2, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1481 |
15634 | 1482 "movq "DCTSIZE_S"*4*2(%%"REG_D"), %%mm7 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1483 "paddw %%mm2, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1484 |
15634 | 1485 "movq %%mm5, "DCTSIZE_S"*5*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1486 "paddw %%mm0, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1487 |
15634 | 1488 "movq %%mm6, "DCTSIZE_S"*3*2(%%"REG_D") \n\t" |
1489 | |
1490 "movq %%mm7, "DCTSIZE_S"*4*2(%%"REG_D") \n\t" | |
1491 "add $24, %%"REG_D" \n\t" | |
1492 "sub $2, %%"REG_c" \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1493 "jnz 1b \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1494 "jmp 5f \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1495 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1496 "3: \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1497 //--- non DC2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1498 //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
|
1499 //psraw mm5, 2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1500 //psraw mm0, 2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1501 //psraw mm6, 2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1502 "movq %%mm5, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1503 "psubw %%mm1, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1504 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1505 "psllw $1, %%mm5 \n\t" //'z10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1506 "paddw %%mm1, %%mm3 \n\t" //'z13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1507 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1508 "movq %%mm0, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1509 "psubw %%mm6, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1510 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1511 "movq %%mm5, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1512 "psllw $1, %%mm0 \n\t" //'z12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1513 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1514 "pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm1 \n\t" //- |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1515 "paddw %%mm0, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1516 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1517 "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
|
1518 "paddw %%mm6, %%mm2 \n\t" //'z11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1519 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1520 "pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1521 "movq %%mm2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1522 |
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 "movq "MANGLE(temps)"+0*8, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1525 "psubw %%mm3, %%mm2 \n\t" |
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 "psllw $1, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1528 "paddw %%mm3, %%mm7 \n\t" //'t7 |
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 "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
|
1531 "movq %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1532 //paddw mm7, MM_2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1533 "psraw $2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1534 |
15634 | 1535 "paddw "DCTSIZE_S"*0*2(%%"REG_D"), %%mm4 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1536 "psubw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1537 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1538 "movq "MANGLE(temps)"+1*8, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1539 "paddw %%mm7, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1540 |
15634 | 1541 "movq %%mm6, "DCTSIZE_S"*7*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1542 "paddw %%mm5, %%mm1 \n\t" //'t12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1543 |
15634 | 1544 "movq %%mm4, "DCTSIZE_S"*0*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1545 "psubw %%mm7, %%mm1 \n\t" //'t6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1546 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1547 "movq "MANGLE(temps)"+2*8, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1548 "psubw %%mm5, %%mm0 \n\t" //'t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1549 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1550 "movq "MANGLE(temps)"+3*8, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1551 "movq %%mm3, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1552 |
15634 | 1553 "paddw "DCTSIZE_S"*1*2(%%"REG_D"), %%mm3 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1554 "psubw %%mm1, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1555 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1556 "psubw %%mm1, %%mm2 \n\t" //'t5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1557 "paddw %%mm1, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1558 |
15634 | 1559 "movq %%mm5, "DCTSIZE_S"*6*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1560 "movq %%mm7, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1561 |
15634 | 1562 "paddw "DCTSIZE_S"*2*2(%%"REG_D"), %%mm7 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1563 "psubw %%mm2, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1564 |
15634 | 1565 "paddw "DCTSIZE_S"*5*2(%%"REG_D"), %%mm4 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1566 "paddw %%mm2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1567 |
15634 | 1568 "movq %%mm3, "DCTSIZE_S"*1*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1569 "paddw %%mm2, %%mm0 \n\t" //'t4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1570 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1571 // 't4 't6 't5 - - - - 't7 |
15634 | 1572 "movq %%mm7, "DCTSIZE_S"*2*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1573 "movq %%mm6, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1574 |
15634 | 1575 "paddw "DCTSIZE_S"*4*2(%%"REG_D"), %%mm6 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1576 "psubw %%mm0, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1577 |
15634 | 1578 "paddw "DCTSIZE_S"*3*2(%%"REG_D"), %%mm1 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1579 "paddw %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1580 |
15634 | 1581 "movq %%mm4, "DCTSIZE_S"*5*2(%%"REG_D") \n\t" |
1582 "add $24, %%"REG_S" \n\t" | |
1583 | |
1584 "movq %%mm6, "DCTSIZE_S"*4*2(%%"REG_D") \n\t" | |
1585 | |
1586 "movq %%mm1, "DCTSIZE_S"*3*2(%%"REG_D") \n\t" | |
1587 "add $24, %%"REG_D" \n\t" | |
1588 "sub $2, %%"REG_c" \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1589 "jnz 1b \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1590 "5: \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1591 |
15632 | 1592 : "+S"(data), "+D"(output), "+c"(cnt)// input regs |
1593 : "d"(thr_adr) | |
15634 | 1594 : "%"REG_a |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1595 ); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1596 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1597 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1598 #endif // HAVE_MMX |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1599 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1600 #ifndef HAVE_MMX |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1601 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1602 static void row_idct_c(DCTELEM* workspace, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1603 int16_t* output_adr, int output_stride, int cnt) |
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 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
|
1606 int_simd16_t tmp10, tmp11, tmp12, tmp13; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1607 int_simd16_t z5, z10, z11, z12, z13; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1608 int16_t* outptr; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1609 DCTELEM* wsptr; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1610 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1611 cnt*=4; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1612 wsptr = workspace; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1613 outptr = output_adr; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1614 for (; cnt > 0; cnt--) { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1615 // Even part |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1616 //Simd version reads 4x4 block and transposes it |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1617 tmp10 = ( wsptr[2] + wsptr[3]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1618 tmp11 = ( wsptr[2] - wsptr[3]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1619 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1620 tmp13 = ( wsptr[0] + wsptr[1]); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1621 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
|
1622 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1623 tmp0 = tmp10 + tmp13; //->temps |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1624 tmp3 = tmp10 - tmp13; //->temps |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1625 tmp1 = tmp11 + tmp12; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1626 tmp2 = tmp11 - tmp12; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1627 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1628 // Odd part |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1629 //Also transpose, with previous: |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1630 // ---- ---- |||| |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1631 // ---- ---- idct |||| |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1632 // ---- ---- ---> |||| |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1633 // ---- ---- |||| |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1634 z13 = wsptr[4] + wsptr[5]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1635 z10 = wsptr[4] - wsptr[5]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1636 z11 = wsptr[6] + wsptr[7]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1637 z12 = wsptr[6] - wsptr[7]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1638 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1639 tmp7 = z11 + z13; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1640 tmp11 = MULTIPLY16H(z11 - z13, FIX_1_414213562); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1641 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1642 z5 = MULTIPLY16H(z10 + z12, FIX_1_847759065); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1643 tmp10 = MULTIPLY16H(z12, FIX_1_082392200) - z5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1644 tmp12 = MULTIPLY16H(z10, FIX_2_613125930) + z5; // - FIX_ |
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 tmp6 = (tmp12<<3) - tmp7; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1647 tmp5 = (tmp11<<3) - tmp6; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1648 tmp4 = (tmp10<<3) + tmp5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1649 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1650 // Final output stage: descale and write column |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1651 outptr[0*output_stride]+= DESCALE(tmp0 + tmp7, 3); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1652 outptr[1*output_stride]+= DESCALE(tmp1 + tmp6, 3); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1653 outptr[2*output_stride]+= DESCALE(tmp2 + tmp5, 3); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1654 outptr[3*output_stride]+= DESCALE(tmp3 - tmp4, 3); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1655 outptr[4*output_stride]+= DESCALE(tmp3 + tmp4, 3); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1656 outptr[5*output_stride]+= DESCALE(tmp2 - tmp5, 3); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1657 outptr[6*output_stride]+= DESCALE(tmp1 - tmp6, 3); //no += ? |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1658 outptr[7*output_stride]+= DESCALE(tmp0 - tmp7, 3); //no += ? |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1659 outptr++; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1660 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1661 wsptr += DCTSIZE; // advance pointer to next row |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1662 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1663 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1664 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1665 #else /* HAVE_MMX */ |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1666 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1667 static void row_idct_mmx (DCTELEM* workspace, |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1668 int16_t* output_adr, int output_stride, int cnt) |
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 asm volatile( |
15634 | 1671 "lea (%%"REG_a",%%"REG_a",2), %%"REG_d" \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1672 "1: \n\t" |
15634 | 1673 "movq "DCTSIZE_S"*0*2(%%"REG_S"), %%mm0 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1674 // |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1675 |
15634 | 1676 "movq "DCTSIZE_S"*1*2(%%"REG_S"), %%mm1 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1677 "movq %%mm0, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1678 |
15634 | 1679 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1680 "punpcklwd %%mm1, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1681 |
15634 | 1682 "movq "DCTSIZE_S"*3*2(%%"REG_S"), %%mm3 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1683 "punpckhwd %%mm1, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1684 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1685 //transpose 4x4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1686 "movq %%mm2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1687 "punpcklwd %%mm3, %%mm2 \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 %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1690 "punpckldq %%mm2, %%mm0 \n\t" //0 |
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 "punpckhdq %%mm2, %%mm6 \n\t" //1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1693 "movq %%mm0, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1694 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1695 "punpckhwd %%mm3, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1696 "psubw %%mm6, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1697 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1698 "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
|
1699 "movq %%mm4, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1700 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1701 "punpckldq %%mm7, %%mm4 \n\t" //2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1702 "paddw %%mm6, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1703 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1704 "punpckhdq %%mm7, %%mm2 \n\t" //3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1705 "movq %%mm4, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1706 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1707 "psllw $2, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1708 "paddw %%mm2, %%mm4 \n\t" //t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1709 |
15634 | 1710 "movq "DCTSIZE_S"*0*2+"DCTSIZE_S"(%%"REG_S"), %%mm3 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1711 "psubw %%mm2, %%mm1 \n\t" //t11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1712 |
15634 | 1713 "movq "DCTSIZE_S"*1*2+"DCTSIZE_S"(%%"REG_S"), %%mm2 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1714 "psubw %%mm5, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1715 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1716 "movq %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1717 "paddw %%mm5, %%mm4 \n\t" //t0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1718 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1719 "psubw %%mm5, %%mm6 \n\t" //t3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1720 "movq %%mm1, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1721 |
15634 | 1722 "movq "DCTSIZE_S"*2*2+"DCTSIZE_S"(%%"REG_S"), %%mm5 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1723 "paddw %%mm0, %%mm1 \n\t" //t1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1724 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1725 "movq %%mm4, "MANGLE(temps)"+0*8 \n\t" //t0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1726 "movq %%mm3, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1727 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1728 "movq %%mm6, "MANGLE(temps)"+1*8 \n\t" //t3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1729 "punpcklwd %%mm2, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1730 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1731 //transpose 4x4 |
15634 | 1732 "movq "DCTSIZE_S"*3*2+"DCTSIZE_S"(%%"REG_S"), %%mm6 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1733 "punpckhwd %%mm2, %%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 %%mm5, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1736 "punpcklwd %%mm6, %%mm5 \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 "psubw %%mm0, %%mm7 \n\t" //t2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1739 "punpckhwd %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1740 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1741 "movq %%mm3, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1742 "punpckldq %%mm5, %%mm3 \n\t" //4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1743 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1744 "punpckhdq %%mm5, %%mm0 \n\t" //5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1745 "movq %%mm4, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1746 |
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, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1749 "punpckldq %%mm2, %%mm4 \n\t" //6 |
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 "psubw %%mm0, %%mm3 \n\t" //z10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1752 "punpckhdq %%mm2, %%mm5 \n\t" //7 |
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 "paddw %%mm0, %%mm6 \n\t" //z13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1755 "movq %%mm4, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1756 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1757 "movq %%mm3, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1758 "psubw %%mm5, %%mm4 \n\t" //z12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1759 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1760 "pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm0 \n\t" //- |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1761 "paddw %%mm4, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1762 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1763 "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
|
1764 "paddw %%mm5, %%mm2 \n\t" //z11 > |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1765 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1766 "pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1767 "movq %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1768 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1769 "psubw %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1770 "paddw %%mm6, %%mm5 \n\t" //t7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1771 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1772 "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
|
1773 "paddw %%mm3, %%mm0 \n\t" //t12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1774 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1775 "psllw $3, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1776 "psubw %%mm3, %%mm4 \n\t" //t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1777 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1778 "movq "MANGLE(temps)"+0*8, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1779 "movq %%mm1, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1780 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1781 "psllw $3, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1782 "psubw %%mm5, %%mm0 \n\t" //t6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1783 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1784 "psllw $3, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1785 "paddw %%mm0, %%mm1 \n\t" //d1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1786 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1787 "psubw %%mm0, %%mm2 \n\t" //t5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1788 "psubw %%mm0, %%mm3 \n\t" //d6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1789 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1790 "paddw %%mm2, %%mm4 \n\t" //t4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1791 "movq %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1792 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1793 "paddw %%mm2, %%mm7 \n\t" //d2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1794 "psubw %%mm2, %%mm0 \n\t" //d5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1795 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1796 "movq "MANGLE(MM_DESCALE_RND)", %%mm2 \n\t" //4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1797 "psubw %%mm5, %%mm6 \n\t" //d7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1798 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1799 "paddw "MANGLE(temps)"+0*8, %%mm5 \n\t" //d0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1800 "paddw %%mm2, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1801 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1802 "paddw %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1803 "psraw $3, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1804 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1805 "paddw %%mm2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1806 "psraw $3, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1807 |
15634 | 1808 "paddw (%%"REG_D"), %%mm5 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1809 "psraw $3, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1810 |
15634 | 1811 "paddw (%%"REG_D",%%"REG_a",), %%mm1 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1812 "paddw %%mm2, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1813 |
15634 | 1814 "paddw (%%"REG_D",%%"REG_a",2), %%mm7 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1815 "paddw %%mm2, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1816 |
15634 | 1817 "movq %%mm5, (%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1818 "paddw %%mm2, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1819 |
15634 | 1820 "movq %%mm1, (%%"REG_D",%%"REG_a",) \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1821 "psraw $3, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1822 |
15634 | 1823 "movq %%mm7, (%%"REG_D",%%"REG_a",2) \n\t" |
1824 "add %%"REG_d", %%"REG_D" \n\t" //3*ls | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1825 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1826 "movq "MANGLE(temps)"+1*8, %%mm5 \n\t" //t3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1827 "psraw $3, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1828 |
15634 | 1829 "paddw (%%"REG_D",%%"REG_a",2), %%mm0 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1830 "psubw %%mm4, %%mm5 \n\t" //d3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1831 |
15634 | 1832 "paddw (%%"REG_D",%%"REG_d",), %%mm3 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1833 "psraw $3, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1834 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1835 "paddw "MANGLE(temps)"+1*8, %%mm4 \n\t" //d4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1836 "paddw %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1837 |
15634 | 1838 "paddw (%%"REG_D",%%"REG_a",4), %%mm6 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1839 "paddw %%mm2, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1840 |
15634 | 1841 "movq %%mm0, (%%"REG_D",%%"REG_a",2) \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1842 "psraw $3, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1843 |
15634 | 1844 "paddw (%%"REG_D"), %%mm5 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1845 "psraw $3, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1846 |
15634 | 1847 "paddw (%%"REG_D",%%"REG_a",), %%mm4 \n\t" |
1848 "add $"DCTSIZE_S"*2*4, %%"REG_S" \n\t" //4 rows | |
1849 | |
1850 "movq %%mm3, (%%"REG_D",%%"REG_d",) \n\t" | |
1851 "movq %%mm6, (%%"REG_D",%%"REG_a",4) \n\t" | |
1852 "movq %%mm5, (%%"REG_D") \n\t" | |
1853 "movq %%mm4, (%%"REG_D",%%"REG_a",) \n\t" | |
1854 | |
1855 "sub %%"REG_d", %%"REG_D" \n\t" | |
1856 "add $8, %%"REG_D" \n\t" | |
1857 "dec %%"REG_c" \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1858 "jnz 1b \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1859 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1860 : "+S"(workspace), "+D"(output_adr), "+c"(cnt) //input regs |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1861 : "a"(output_stride*sizeof(short)) |
15634 | 1862 : "%"REG_d |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1863 ); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1864 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1865 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1866 #endif // HAVE_MMX |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1867 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1868 #ifndef HAVE_MMX |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1869 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1870 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
|
1871 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1872 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
|
1873 int_simd16_t tmp10, tmp11, tmp12, tmp13; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1874 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
|
1875 DCTELEM *dataptr; |
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 cnt*=4; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1878 // Pass 1: process rows. |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1879 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1880 dataptr = data; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1881 for (; cnt > 0; cnt--) { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1882 tmp0 = pixels[line_size*0] + pixels[line_size*7]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1883 tmp7 = pixels[line_size*0] - pixels[line_size*7]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1884 tmp1 = pixels[line_size*1] + pixels[line_size*6]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1885 tmp6 = pixels[line_size*1] - pixels[line_size*6]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1886 tmp2 = pixels[line_size*2] + pixels[line_size*5]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1887 tmp5 = pixels[line_size*2] - pixels[line_size*5]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1888 tmp3 = pixels[line_size*3] + pixels[line_size*4]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1889 tmp4 = pixels[line_size*3] - pixels[line_size*4]; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1890 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1891 // Even part |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1892 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1893 tmp10 = tmp0 + tmp3; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1894 tmp13 = tmp0 - tmp3; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1895 tmp11 = tmp1 + tmp2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1896 tmp12 = tmp1 - tmp2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1897 //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
|
1898 //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
|
1899 //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
|
1900 dataptr[2] = tmp10 + tmp11; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1901 dataptr[3] = tmp10 - tmp11; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1902 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1903 z1 = MULTIPLY16H((tmp12 + tmp13)<<2, FIX_0_707106781); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1904 dataptr[0] = tmp13 + z1; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1905 dataptr[1] = tmp13 - z1; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1906 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1907 // Odd part |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1908 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1909 tmp10 = (tmp4 + tmp5) <<2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1910 tmp11 = (tmp5 + tmp6) <<2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1911 tmp12 = (tmp6 + tmp7) <<2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1912 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1913 z5 = MULTIPLY16H(tmp10 - tmp12, FIX_0_382683433); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1914 z2 = MULTIPLY16H(tmp10, FIX_0_541196100) + z5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1915 z4 = MULTIPLY16H(tmp12, FIX_1_306562965) + z5; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1916 z3 = MULTIPLY16H(tmp11, FIX_0_707106781); |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1917 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1918 z11 = tmp7 + z3; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1919 z13 = tmp7 - z3; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1920 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1921 dataptr[4] = z13 + z2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1922 dataptr[5] = z13 - z2; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1923 dataptr[6] = z11 + z4; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1924 dataptr[7] = z11 - z4; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1925 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1926 pixels++; // advance pointer to next column |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1927 dataptr += DCTSIZE; |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1928 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1929 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1930 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1931 #else /* HAVE_MMX */ |
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 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
|
1934 { |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1935 asm volatile( |
15634 | 1936 "lea (%%"REG_a",%%"REG_a",2), %%"REG_d" \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1937 "6: \n\t" |
15634 | 1938 "movd (%%"REG_S"), %%mm0 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1939 "pxor %%mm7, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1940 |
15634 | 1941 "movd (%%"REG_S",%%"REG_a",), %%mm1 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1942 "punpcklbw %%mm7, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1943 |
15634 | 1944 "movd (%%"REG_S",%%"REG_a",2), %%mm2 \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1945 "punpcklbw %%mm7, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1946 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1947 "punpcklbw %%mm7, %%mm2 \n\t" |
15634 | 1948 "add %%"REG_d", %%"REG_S" \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1949 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1950 "movq %%mm0, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1951 // |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1952 |
15634 | 1953 "movd (%%"REG_S",%%"REG_a",4), %%mm3 \n\t" //7 ;prefetch! |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1954 "movq %%mm1, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1955 |
15634 | 1956 "movd (%%"REG_S",%%"REG_d",), %%mm4 \n\t" //6 |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1957 "punpcklbw %%mm7, %%mm3 \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 "psubw %%mm3, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1960 "punpcklbw %%mm7, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1961 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1962 "paddw %%mm3, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1963 "psubw %%mm4, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1964 |
15634 | 1965 "movd (%%"REG_S",%%"REG_a",2), %%mm3 \n\t" //5 |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1966 "paddw %%mm4, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1967 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1968 "movq %%mm5, "MANGLE(temps)"+0*8 \n\t" //t7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1969 "punpcklbw %%mm7, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1970 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1971 "movq %%mm6, "MANGLE(temps)"+1*8 \n\t" //t6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1972 "movq %%mm2, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1973 |
15634 | 1974 "movd (%%"REG_S"), %%mm5 \n\t" //3 |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1975 "paddw %%mm3, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1976 |
15634 | 1977 "movd (%%"REG_S",%%"REG_a",), %%mm6 \n\t" //4 |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1978 "punpcklbw %%mm7, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1979 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1980 "psubw %%mm3, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1981 "punpcklbw %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1982 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1983 "movq %%mm5, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1984 "paddw %%mm6, %%mm5 \n\t" //t3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1985 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1986 "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
|
1987 "movq %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1988 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1989 "movq %%mm1, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1990 "psubw %%mm5, %%mm0 \n\t" //t13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1991 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1992 "psubw %%mm2, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1993 "paddw %%mm2, %%mm7 \n\t" //t11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1994 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1995 "paddw %%mm0, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1996 "movq %%mm7, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1997 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1998 "psllw $2, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
1999 "paddw %%mm5, %%mm6 \n\t" //t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2000 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2001 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2002 "paddw %%mm6, %%mm7 \n\t" //d2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2003 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2004 "psubw %%mm2, %%mm6 \n\t" //d3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2005 "movq %%mm0, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2006 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2007 //transpose 4x4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2008 "movq %%mm7, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2009 "punpcklwd %%mm6, %%mm7 \n\t" |
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 "paddw %%mm1, %%mm0 \n\t" //d0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2012 "punpckhwd %%mm6, %%mm2 \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 "psubw %%mm1, %%mm5 \n\t" //d1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2015 "movq %%mm0, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2016 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2017 "movq "MANGLE(temps)"+1*8, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2018 "punpcklwd %%mm5, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2019 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2020 "punpckhwd %%mm5, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2021 "movq %%mm0, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2022 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2023 "punpckldq %%mm7, %%mm0 \n\t" //0 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2024 "paddw %%mm4, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2025 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2026 "punpckhdq %%mm7, %%mm5 \n\t" //1 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2027 "movq %%mm6, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2028 |
15634 | 2029 "movq %%mm0, "DCTSIZE_S"*0*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2030 "punpckldq %%mm2, %%mm6 \n\t" //2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2031 |
15634 | 2032 "movq %%mm5, "DCTSIZE_S"*1*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2033 "punpckhdq %%mm2, %%mm7 \n\t" //3 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2034 |
15634 | 2035 "movq %%mm6, "DCTSIZE_S"*2*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2036 "paddw %%mm1, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2037 |
15634 | 2038 "movq %%mm7, "DCTSIZE_S"*3*2(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2039 "psllw $2, %%mm3 \n\t" //t10 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2040 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2041 "movq "MANGLE(temps)"+0*8, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2042 "psllw $2, %%mm4 \n\t" //t11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2043 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2044 "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
|
2045 "paddw %%mm2, %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2046 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2047 "psllw $2, %%mm1 \n\t" //t12 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2048 "movq %%mm3, %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2049 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2050 "pmulhw "MANGLE(MM_FIX_0_541196100)", %%mm0 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2051 "psubw %%mm1, %%mm3 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2052 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2053 "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
|
2054 "movq %%mm2, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2055 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2056 "pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm1 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2057 "psubw %%mm4, %%mm2 \n\t" //z13 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2058 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2059 "paddw %%mm4, %%mm5 \n\t" //z11 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2060 "movq %%mm2, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2061 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2062 "paddw %%mm3, %%mm0 \n\t" //z2 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2063 "movq %%mm5, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2064 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2065 "paddw %%mm0, %%mm2 \n\t" //d4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2066 "psubw %%mm0, %%mm6 \n\t" //d5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2067 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2068 "movq %%mm2, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2069 "paddw %%mm3, %%mm1 \n\t" //z4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2070 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2071 //transpose 4x4 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2072 "punpcklwd %%mm6, %%mm2 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2073 "paddw %%mm1, %%mm5 \n\t" //d6 |
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 "punpckhwd %%mm6, %%mm4 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2076 "psubw %%mm1, %%mm7 \n\t" //d7 |
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 "movq %%mm5, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2079 "punpcklwd %%mm7, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2080 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2081 "punpckhwd %%mm7, %%mm6 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2082 "movq %%mm2, %%mm7 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2083 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2084 "punpckldq %%mm5, %%mm2 \n\t" //4 |
15634 | 2085 "sub %%"REG_d", %%"REG_S" \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2086 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2087 "punpckhdq %%mm5, %%mm7 \n\t" //5 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2088 "movq %%mm4, %%mm5 \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2089 |
15634 | 2090 "movq %%mm2, "DCTSIZE_S"*0*2+"DCTSIZE_S"(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2091 "punpckldq %%mm6, %%mm4 \n\t" //6 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2092 |
15634 | 2093 "movq %%mm7, "DCTSIZE_S"*1*2+"DCTSIZE_S"(%%"REG_D") \n\t" |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2094 "punpckhdq %%mm6, %%mm5 \n\t" //7 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2095 |
15634 | 2096 "movq %%mm4, "DCTSIZE_S"*2*2+"DCTSIZE_S"(%%"REG_D") \n\t" |
2097 "add $4, %%"REG_S" \n\t" | |
2098 | |
2099 "movq %%mm5, "DCTSIZE_S"*3*2+"DCTSIZE_S"(%%"REG_D") \n\t" | |
2100 "add $"DCTSIZE_S"*2*4, %%"REG_D" \n\t" //4 rows | |
2101 "dec %%"REG_c" \n\t" | |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2102 "jnz 6b \n\t" |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2103 |
15632 | 2104 : "+S"(pixels), "+D"(data), "+c"(cnt) //input regs |
2105 : "a"(line_size) | |
15634 | 2106 : "%"REG_d); |
15631
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2107 } |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2108 |
d5a95e6f5f07
faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff
changeset
|
2109 #endif // HAVE_MMX |