annotate libmpcodecs/vf_fspp.c @ 35612:488b3fea696d

Allow libmng autodetection to test for "-lmng" and "-lmng -ljpeg". Patch by Josh, joshf87 live com
author cehoyos
date Sat, 22 Dec 2012 17:53:30 +0000
parents a28c4ba97171
children b4ce15212bfc
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1 /*
26727
82601a38e2a7 Use standard license headers.
diego
parents: 26052
diff changeset
2 * Copyright (C) 2003 Michael Niedermayer <michaelni@gmx.at>
82601a38e2a7 Use standard license headers.
diego
parents: 26052
diff changeset
3 * Copyright (C) 2005 Nikolaj Poroshin <porosh3@psu.ru>
82601a38e2a7 Use standard license headers.
diego
parents: 26052
diff changeset
4 *
82601a38e2a7 Use standard license headers.
diego
parents: 26052
diff changeset
5 * This file is part of MPlayer.
82601a38e2a7 Use standard license headers.
diego
parents: 26052
diff changeset
6 *
82601a38e2a7 Use standard license headers.
diego
parents: 26052
diff changeset
7 * MPlayer is free software; you can redistribute it and/or modify
82601a38e2a7 Use standard license headers.
diego
parents: 26052
diff changeset
8 * it under the terms of the GNU General Public License as published by
82601a38e2a7 Use standard license headers.
diego
parents: 26052
diff changeset
9 * the Free Software Foundation; either version 2 of the License, or
82601a38e2a7 Use standard license headers.
diego
parents: 26052
diff changeset
10 * (at your option) any later version.
82601a38e2a7 Use standard license headers.
diego
parents: 26052
diff changeset
11 *
82601a38e2a7 Use standard license headers.
diego
parents: 26052
diff changeset
12 * MPlayer is distributed in the hope that it will be useful,
82601a38e2a7 Use standard license headers.
diego
parents: 26052
diff changeset
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
82601a38e2a7 Use standard license headers.
diego
parents: 26052
diff changeset
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
82601a38e2a7 Use standard license headers.
diego
parents: 26052
diff changeset
15 * GNU General Public License for more details.
82601a38e2a7 Use standard license headers.
diego
parents: 26052
diff changeset
16 *
82601a38e2a7 Use standard license headers.
diego
parents: 26052
diff changeset
17 * You should have received a copy of the GNU General Public License along
82601a38e2a7 Use standard license headers.
diego
parents: 26052
diff changeset
18 * with MPlayer; if not, write to the Free Software Foundation, Inc.,
82601a38e2a7 Use standard license headers.
diego
parents: 26052
diff changeset
19 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
82601a38e2a7 Use standard license headers.
diego
parents: 26052
diff changeset
20 */
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
21
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
22 /*
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
23 * This implementation is based on an algorithm described in
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
24 * "Aria Nosratinia Embedded Post-Processing for
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
25 * Enhancement of Compressed Images (1999)"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
26 * (http://citeseer.nj.nec.com/nosratinia99embedded.html)
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
27 * 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
28 * 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
29 */
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
30
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 Heavily optimized version of SPP filter by Nikolaj
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
33 */
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
34
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
35 #include <stdio.h>
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
36 #include <stdlib.h>
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
37 #include <string.h>
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
38 #include <inttypes.h>
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
39 #include <math.h>
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
40
17012
6ff3379a0862 Unify include path handling, -I.. is in CFLAGS.
diego
parents: 16018
diff changeset
41 #include "config.h"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
42
17012
6ff3379a0862 Unify include path handling, -I.. is in CFLAGS.
diego
parents: 16018
diff changeset
43 #include "mp_msg.h"
6ff3379a0862 Unify include path handling, -I.. is in CFLAGS.
diego
parents: 16018
diff changeset
44 #include "cpudetect.h"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
45 #include "img_format.h"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
46 #include "mp_image.h"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
47 #include "vf.h"
33871
30f5e5cd3676 Move code for setting up libav* logging callbacks from vd_ffmpeg to a
reimar
parents: 32702
diff changeset
48 #include "av_helpers.h"
17012
6ff3379a0862 Unify include path handling, -I.. is in CFLAGS.
diego
parents: 16018
diff changeset
49 #include "libvo/fastmemcpy.h"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
50
28588
7f03a6d3c941 Move FFmpeg #includes below all others so that they do not override
diego
parents: 28327
diff changeset
51 #include "libavutil/internal.h"
7f03a6d3c941 Move FFmpeg #includes below all others so that they do not override
diego
parents: 28327
diff changeset
52 #include "libavutil/intreadwrite.h"
7f03a6d3c941 Move FFmpeg #includes below all others so that they do not override
diego
parents: 28327
diff changeset
53 #include "libavutil/mem.h"
34198
064be070d8c8 cpudetect: clean up libavutil #includes
diego
parents: 33871
diff changeset
54 #include "libavutil/x86_cpu.h"
28588
7f03a6d3c941 Move FFmpeg #includes below all others so that they do not override
diego
parents: 28327
diff changeset
55 #include "libavcodec/avcodec.h"
7f03a6d3c941 Move FFmpeg #includes below all others so that they do not override
diego
parents: 28327
diff changeset
56 #include "libavcodec/dsputil.h"
7f03a6d3c941 Move FFmpeg #includes below all others so that they do not override
diego
parents: 28327
diff changeset
57
28327
c39a1fd7d45c Fix compilation after DECLARE_ASM_CONST/DECLARE_ALIGNED moving within FFmpeg.
diego
parents: 28290
diff changeset
58 #undef free
c39a1fd7d45c Fix compilation after DECLARE_ASM_CONST/DECLARE_ALIGNED moving within FFmpeg.
diego
parents: 28290
diff changeset
59 #undef malloc
c39a1fd7d45c Fix compilation after DECLARE_ASM_CONST/DECLARE_ALIGNED moving within FFmpeg.
diego
parents: 28290
diff changeset
60
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
61 //===========================================================================//
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
62 #define BLOCKSZ 12
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
63
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
64 static const short custom_threshold[64]=
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
65 // values (296) can't be too high
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
66 // -it causes too big quant dependence
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
67 // or maybe overflow(check), which results in some flashing
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
68 { 71, 296, 295, 237, 71, 40, 38, 19,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
69 245, 193, 185, 121, 102, 73, 53, 27,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
70 158, 129, 141, 107, 97, 73, 50, 26,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
71 102, 116, 109, 98, 82, 66, 45, 23,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
72 71, 94, 95, 81, 70, 56, 38, 20,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
73 56, 77, 74, 66, 56, 44, 30, 15,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
74 38, 53, 50, 45, 38, 30, 21, 11,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
75 20, 27, 26, 23, 20, 15, 11, 5
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
76 };
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
77
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
78 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
79 { 0, 48, 12, 60, 3, 51, 15, 63, },
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
80 { 32, 16, 44, 28, 35, 19, 47, 31, },
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
81 { 8, 56, 4, 52, 11, 59, 7, 55, },
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
82 { 40, 24, 36, 20, 43, 27, 39, 23, },
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
83 { 2, 50, 14, 62, 1, 49, 13, 61, },
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
84 { 34, 18, 46, 30, 33, 17, 45, 29, },
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
85 { 10, 58, 6, 54, 9, 57, 5, 53, },
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
86 { 42, 26, 38, 22, 41, 25, 37, 21, },
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
87 };
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
88
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
89 struct vf_priv_s { //align 16 !
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
90 uint64_t threshold_mtx_noq[8*2];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
91 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
92
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
93 int log2_count;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
94 int temp_stride;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
95 int qp;
17225
ec9888363742 reverse the H264 hack
henry
parents: 17223
diff changeset
96 int mpeg2;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
97 int prev_q;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
98 uint8_t *src;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
99 int16_t *temp;
17225
ec9888363742 reverse the H264 hack
henry
parents: 17223
diff changeset
100 int bframes;
17133
a2b24e0d7772 prevent flicker on b-frames, trivial port from vf_spp
henry
parents: 17012
diff changeset
101 char *non_b_qp;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
102 };
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
28290
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_...
reimar
parents: 27754
diff changeset
105 #if !HAVE_MMX
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
106
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
107 //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
108 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
109 {int y, x;
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
110 #define STORE(pos) \
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
111 temp= (src[x + pos] + (d[pos]>>log2_scale))>>(6-log2_scale); \
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
112 src[x + pos]=src[x + pos - 8*src_stride]=0; \
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
113 if(temp & 0x100) temp= ~(temp>>31); \
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
114 dst[x + pos]= temp;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
115
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
116 for(y=0; y<height; y++){
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
117 const uint8_t *d= dither[y];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
118 for(x=0; x<width; x+=8){
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
119 int temp;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
120 STORE(0);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
121 STORE(1);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
122 STORE(2);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
123 STORE(3);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
124 STORE(4);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
125 STORE(5);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
126 STORE(6);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
127 STORE(7);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
128 }
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
129 src+=src_stride;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
130 dst+=dst_stride;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
131 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
132 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
133
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
134 //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
135 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
136 {int y, x;
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
137 #define STORE2(pos) \
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
138 temp= (src[x + pos] + src[x + pos + 16*src_stride] + (d[pos]>>log2_scale))>>(6-log2_scale); \
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
139 src[x + pos + 16*src_stride]=0; \
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
140 if(temp & 0x100) temp= ~(temp>>31); \
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
141 dst[x + pos]= temp;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
142
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
143 for(y=0; y<height; y++){
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
144 const uint8_t *d= dither[y];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
145 for(x=0; x<width; x+=8){
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
146 int temp;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
147 STORE2(0);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
148 STORE2(1);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
149 STORE2(2);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
150 STORE2(3);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
151 STORE2(4);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
152 STORE2(5);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
153 STORE2(6);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
154 STORE2(7);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
155 }
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
156 src+=src_stride;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
157 dst+=dst_stride;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
158 }
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
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
161 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
162 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
163 int a;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
164 for(a=0;a<64;a++)
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
165 ((short*)p->threshold_mtx)[a]=q * ((short*)p->threshold_mtx_noq)[a];//ints faster in C
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
166 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
167
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
168 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
169 static void row_idct_c(DCTELEM* workspace,
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
170 int16_t* output_adr, int output_stride, int cnt);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
171 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
172
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
173 //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
174 #define store_slice_s store_slice_c
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
175 #define store_slice2_s store_slice2_c
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
176 #define mul_thrmat_s mul_thrmat_c
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
177 #define column_fidct_s column_fidct_c
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
178 #define row_idct_s row_idct_c
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
179 #define row_fdct_s row_fdct_c
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
180
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
181 #else /* HAVE_MMX */
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 //This func reads from 1 slice, 1 and clears 0 & 1
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
184 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
185 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
186 const uint8_t *od=&dither[0][0];
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
187 const uint8_t *end=&dither[height][0];
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
188 width = (width+7)&~7;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
189 dst_stride-=width;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
190 //src_stride=(src_stride-width)*2;
27754
08d18fe9da52 Change all occurrences of asm and __asm to __asm__, same as was done for FFmpeg.
diego
parents: 26727
diff changeset
191 __asm__ volatile(
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
192 "mov %5, %%"REG_d" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
193 "mov %6, %%"REG_S" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
194 "mov %7, %%"REG_D" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
195 "mov %1, %%"REG_a" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
196 "movd %%"REG_d", %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
197 "xor $-1, %%"REG_d" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
198 "mov %%"REG_a", %%"REG_c" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
199 "add $7, %%"REG_d" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
200 "neg %%"REG_a" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
201 "sub %0, %%"REG_c" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
202 "add %%"REG_c", %%"REG_c" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
203 "movd %%"REG_d", %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
204 "mov %%"REG_c", %1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
205 "mov %2, %%"REG_d" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
206 "shl $4, %%"REG_a" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
207
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
208 "2: \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
209 "movq (%%"REG_d"), %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
210 "movq %%mm3, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
211 "pxor %%mm7, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
212 "punpcklbw %%mm7, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
213 "punpckhbw %%mm7, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
214 "mov %0, %%"REG_c" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
215 "psraw %%mm5, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
216 "psraw %%mm5, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
217 "1: \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
218 "movq %%mm7, (%%"REG_S",%%"REG_a",) \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
219 "movq (%%"REG_S"), %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
220 "movq 8(%%"REG_S"), %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
221
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
222 "movq %%mm7, 8(%%"REG_S",%%"REG_a",) \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
223 "paddw %%mm3, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
224 "paddw %%mm4, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
225
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
226 "movq %%mm7, (%%"REG_S") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
227 "psraw %%mm2, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
228 "psraw %%mm2, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
229
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
230 "movq %%mm7, 8(%%"REG_S") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
231 "packuswb %%mm1, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
232 "add $16, %%"REG_S" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
233
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
234 "movq %%mm0, (%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
235 "add $8, %%"REG_D" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
236 "sub $8, %%"REG_c" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
237 "jg 1b \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
238 "add %1, %%"REG_S" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
239 "add $8, %%"REG_d" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
240 "add %3, %%"REG_D" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
241 "cmp %4, %%"REG_d" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
242 "jl 2b \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
243
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
244 :
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
245 : "m" (width), "m" (src_stride), "erm" (od), "m" (dst_stride), "erm" (end),
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
246 "m" (log2_scale), "m" (src), "m" (dst) //input
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
247 : "%"REG_a, "%"REG_c, "%"REG_d, "%"REG_S, "%"REG_D
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
248 );
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
249 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
250
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
251 //This func reads from 2 slices, 0 & 2 and clears 2-nd
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
252 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
253 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
254 const uint8_t *od=&dither[0][0];
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
255 const uint8_t *end=&dither[height][0];
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
256 width = (width+7)&~7;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
257 dst_stride-=width;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
258 //src_stride=(src_stride-width)*2;
27754
08d18fe9da52 Change all occurrences of asm and __asm to __asm__, same as was done for FFmpeg.
diego
parents: 26727
diff changeset
259 __asm__ volatile(
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
260 "mov %5, %%"REG_d" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
261 "mov %6, %%"REG_S" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
262 "mov %7, %%"REG_D" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
263 "mov %1, %%"REG_a" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
264 "movd %%"REG_d", %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
265 "xor $-1, %%"REG_d" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
266 "mov %%"REG_a", %%"REG_c" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
267 "add $7, %%"REG_d" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
268 "sub %0, %%"REG_c" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
269 "add %%"REG_c", %%"REG_c" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
270 "movd %%"REG_d", %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
271 "mov %%"REG_c", %1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
272 "mov %2, %%"REG_d" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
273 "shl $5, %%"REG_a" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
274
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
275 "2: \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
276 "movq (%%"REG_d"), %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
277 "movq %%mm3, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
278 "pxor %%mm7, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
279 "punpcklbw %%mm7, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
280 "punpckhbw %%mm7, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
281 "mov %0, %%"REG_c" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
282 "psraw %%mm5, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
283 "psraw %%mm5, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
284 "1: \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
285 "movq (%%"REG_S"), %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
286 "movq 8(%%"REG_S"), %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
287 "paddw %%mm3, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
288
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
289 "paddw (%%"REG_S",%%"REG_a",), %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
290 "paddw %%mm4, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
291 "movq 8(%%"REG_S",%%"REG_a",), %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
292
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
293 "movq %%mm7, (%%"REG_S",%%"REG_a",) \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
294 "psraw %%mm2, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
295 "paddw %%mm6, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
296
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
297 "movq %%mm7, 8(%%"REG_S",%%"REG_a",) \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
298 "psraw %%mm2, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
299 "packuswb %%mm1, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
300
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
301 "movq %%mm0, (%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
302 "add $16, %%"REG_S" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
303 "add $8, %%"REG_D" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
304 "sub $8, %%"REG_c" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
305 "jg 1b \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
306 "add %1, %%"REG_S" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
307 "add $8, %%"REG_d" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
308 "add %3, %%"REG_D" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
309 "cmp %4, %%"REG_d" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
310 "jl 2b \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
311
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
312 :
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
313 : "m" (width), "m" (src_stride), "erm" (od), "m" (dst_stride), "erm" (end),
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
314 "m" (log2_scale), "m" (src), "m" (dst) //input
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
315 : "%"REG_a, "%"REG_c, "%"REG_d, "%"REG_D, "%"REG_S
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
316 );
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
317 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
318
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
319 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
320 {
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
321 uint64_t *adr=&p->threshold_mtx_noq[0];
27754
08d18fe9da52 Change all occurrences of asm and __asm to __asm__, same as was done for FFmpeg.
diego
parents: 26727
diff changeset
322 __asm__ volatile(
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
323 "movd %0, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
324 "add $8*8*2, %%"REG_D" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
325 "movq 0*8(%%"REG_S"), %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
326 "punpcklwd %%mm7, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
327 "movq 1*8(%%"REG_S"), %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
328 "punpckldq %%mm7, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
329 "pmullw %%mm7, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
330
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
331 "movq 2*8(%%"REG_S"), %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
332 "pmullw %%mm7, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
333
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
334 "movq 3*8(%%"REG_S"), %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
335 "pmullw %%mm7, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
336
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
337 "movq %%mm0, 0*8(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
338 "movq 4*8(%%"REG_S"), %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
339 "pmullw %%mm7, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
340
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
341 "movq %%mm1, 1*8(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
342 "movq 5*8(%%"REG_S"), %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
343 "pmullw %%mm7, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
344
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
345 "movq %%mm2, 2*8(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
346 "movq 6*8(%%"REG_S"), %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
347 "pmullw %%mm7, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
348
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
349 "movq %%mm3, 3*8(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
350 "movq 7*8+0*8(%%"REG_S"), %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
351 "pmullw %%mm7, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
352
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
353 "movq %%mm4, 4*8(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
354 "movq 7*8+1*8(%%"REG_S"), %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
355 "pmullw %%mm7, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
356
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
357 "movq %%mm5, 5*8(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
358 "movq 7*8+2*8(%%"REG_S"), %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
359 "pmullw %%mm7, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
360
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
361 "movq %%mm6, 6*8(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
362 "movq 7*8+3*8(%%"REG_S"), %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
363 "pmullw %%mm7, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
364
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
365 "movq %%mm0, 7*8+0*8(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
366 "movq 7*8+4*8(%%"REG_S"), %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
367 "pmullw %%mm7, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
368
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
369 "movq %%mm1, 7*8+1*8(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
370 "movq 7*8+5*8(%%"REG_S"), %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
371 "pmullw %%mm7, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
372
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
373 "movq %%mm2, 7*8+2*8(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
374 "movq 7*8+6*8(%%"REG_S"), %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
375 "pmullw %%mm7, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
376
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
377 "movq %%mm3, 7*8+3*8(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
378 "movq 14*8+0*8(%%"REG_S"), %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
379 "pmullw %%mm7, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
380
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
381 "movq %%mm4, 7*8+4*8(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
382 "movq 14*8+1*8(%%"REG_S"), %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
383 "pmullw %%mm7, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
384
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
385 "movq %%mm5, 7*8+5*8(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
386 "pmullw %%mm7, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
387
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
388 "movq %%mm6, 7*8+6*8(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
389 "movq %%mm0, 14*8+0*8(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
390 "movq %%mm1, 14*8+1*8(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
391
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
392 : "+g" (q), "+S" (adr), "+D" (adr)
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
393 :
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
394 );
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
395 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
396
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
397 static void column_fidct_mmx(int16_t* thr_adr, DCTELEM *data, DCTELEM *output, int cnt);
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
398 static void row_idct_mmx(DCTELEM* workspace,
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
399 int16_t* output_adr, int output_stride, int cnt);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
400 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
401
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
402 #define store_slice_s store_slice_mmx
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
403 #define store_slice2_s store_slice2_mmx
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
404 #define mul_thrmat_s mul_thrmat_mmx
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
405 #define column_fidct_s column_fidct_mmx
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
406 #define row_idct_s row_idct_mmx
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
407 #define row_fdct_s row_fdct_mmx
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
408 #endif // HAVE_MMX
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
409
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
410 static void filter(struct vf_priv_s *p, uint8_t *dst, uint8_t *src,
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
411 int dst_stride, int src_stride,
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
412 int width, int height,
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
413 uint8_t *qp_store, int qp_stride, int is_luma)
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
414 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
415 int x, x0, y, es, qy, t;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
416 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
417 const int step=6-p->log2_count;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
418 const int qps= 3 + is_luma;
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
419 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
420 DCTELEM *block= (DCTELEM *)block_align;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
421 DCTELEM *block3=(DCTELEM *)(block_align+4*8*BLOCKSZ);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
422
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
423 memset(block3, 0, 4*8*BLOCKSZ);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
424
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
425 //p->src=src-src_stride*8-8;//!
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
426 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
427 for(y=0; y<height; y++){
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
428 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
429 fast_memcpy(p->src + index, src + y*src_stride, width);//this line can be avoided by using DR & user fr.buffers
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
430 for(x=0; x<8; x++){
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
431 p->src[index - x - 1]= p->src[index + x ];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
432 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
433 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
434 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
435 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
436 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
437 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
438 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
439 //FIXME (try edge emu)
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=8; y<24; y++)
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
442 memset(p->temp+ 8 +y*stride, 0,width*sizeof(int16_t));
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
443
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
444 for(y=step; y<height+8; y+=step){ //step= 1,2
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
445 qy=y-4;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
446 if (qy>height-1) qy=height-1;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
447 if (qy<0) qy=0;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
448 qy=(qy>>qps)*qp_stride;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
449 row_fdct_s(block, p->src + y*stride +2-(y&1), stride, 2);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
450 for(x0=0; x0<width+8-8*(BLOCKSZ-1); x0+=8*(BLOCKSZ-1)){
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
451 row_fdct_s(block+8*8, p->src + y*stride+8+x0 +2-(y&1), stride, 2*(BLOCKSZ-1));
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
452 if(p->qp)
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
453 column_fidct_s((int16_t*)(&p->threshold_mtx[0]), block+0*8, block3+0*8, 8*(BLOCKSZ-1)); //yes, this is a HOTSPOT
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
454 else
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
455 for (x=0; x<8*(BLOCKSZ-1); x+=8) {
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
456 t=x+x0-2; //correct t=x+x0-2-(y&1), but its the same
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
457 if (t<0) t=0;//t always < width-2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
458 t=qp_store[qy+(t>>qps)];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
459 t=norm_qscale(t, p->mpeg2);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
460 if (t!=p->prev_q) p->prev_q=t, mul_thrmat_s(p, t);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
461 column_fidct_s((int16_t*)(&p->threshold_mtx[0]), block+x*8, block3+x*8, 8); //yes, this is a HOTSPOT
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
462 }
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
463 row_idct_s(block3+0*8, p->temp + (y&15)*stride+x0+2-(y&1), stride, 2*(BLOCKSZ-1));
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
464 memmove(block, block+(BLOCKSZ-1)*64, 8*8*sizeof(DCTELEM)); //cycling
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
465 memmove(block3, block3+(BLOCKSZ-1)*64, 6*8*sizeof(DCTELEM));
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
466 }
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
467 //
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
468 es=width+8-x0; // 8, ...
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
469 if (es>8)
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
470 row_fdct_s(block+8*8, p->src + y*stride+8+x0 +2-(y&1), stride, (es-4)>>2);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
471 column_fidct_s((int16_t*)(&p->threshold_mtx[0]), block, block3, es&(~1));
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
472 row_idct_s(block3+0*8, p->temp + (y&15)*stride+x0+2-(y&1), stride, es>>2);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
473 {const int y1=y-8+step;//l5-7 l4-6
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
474 if (!(y1&7) && y1) {
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
475 if (y1&8) store_slice_s(dst + (y1-8)*dst_stride, p->temp+ 8 +8*stride,
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
476 dst_stride, stride, width, 8, 5-p->log2_count);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
477 else store_slice2_s(dst + (y1-8)*dst_stride, p->temp+ 8 +0*stride,
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
478 dst_stride, stride, width, 8, 5-p->log2_count);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
479 } }
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
480 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
481
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
482 if (y&7) { // == height & 7
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
483 if (y&8) store_slice_s(dst + ((y-8)&~7)*dst_stride, p->temp+ 8 +8*stride,
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
484 dst_stride, stride, width, y&7, 5-p->log2_count);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
485 else store_slice2_s(dst + ((y-8)&~7)*dst_stride, p->temp+ 8 +0*stride,
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
486 dst_stride, stride, width, y&7, 5-p->log2_count);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
487 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
488 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
489
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
490 static int config(struct vf_instance *vf,
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
491 int width, int height, int d_width, int d_height,
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
492 unsigned int flags, unsigned int outfmt)
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
493 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
494 int h= (height+16+15)&(~15);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
495
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
496 vf->priv->temp_stride= (width+16+15)&(~15);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
497 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
498 //this can also be avoided, see above
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
499 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
500
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
501 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
502 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
503
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
504 static void get_image(struct vf_instance *vf, mp_image_t *mpi)
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
505 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
506 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
507 // 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
508 vf->dmpi=vf_get_image(vf->next,mpi->imgfmt,
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
509 mpi->type, mpi->flags, mpi->width, mpi->height);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
510 mpi->planes[0]=vf->dmpi->planes[0];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
511 mpi->stride[0]=vf->dmpi->stride[0];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
512 mpi->width=vf->dmpi->width;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
513 if(mpi->flags&MP_IMGFLAG_PLANAR){
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
514 mpi->planes[1]=vf->dmpi->planes[1];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
515 mpi->planes[2]=vf->dmpi->planes[2];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
516 mpi->stride[1]=vf->dmpi->stride[1];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
517 mpi->stride[2]=vf->dmpi->stride[2];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
518 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
519 mpi->flags|=MP_IMGFLAG_DIRECT;
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
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
522 static int put_image(struct vf_instance *vf, mp_image_t *mpi, double pts)
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
523 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
524 mp_image_t *dmpi;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
525 if(!(mpi->flags&MP_IMGFLAG_DIRECT)){
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
526 // no DR, so get a new image! hope we'll get DR buffer:
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
527 dmpi=vf_get_image(vf->next,mpi->imgfmt,
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
528 MP_IMGTYPE_TEMP,
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
529 MP_IMGFLAG_ACCEPT_STRIDE|MP_IMGFLAG_PREFER_ALIGNED_STRIDE,
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
530 mpi->width,mpi->height);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
531 vf_clone_mpi_attributes(dmpi, mpi);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
532 }else{
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
533 dmpi=vf->dmpi;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
534 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
535
17225
ec9888363742 reverse the H264 hack
henry
parents: 17223
diff changeset
536 vf->priv->mpeg2= mpi->qscale_type;
17133
a2b24e0d7772 prevent flicker on b-frames, trivial port from vf_spp
henry
parents: 17012
diff changeset
537 if(mpi->pict_type != 3 && mpi->qscale && !vf->priv->qp){
30363
915be5c7a30c Make sure that a qstride of 0 (intentional or not) does not completely break
reimar
parents: 29310
diff changeset
538 int w = mpi->qstride;
915be5c7a30c Make sure that a qstride of 0 (intentional or not) does not completely break
reimar
parents: 29310
diff changeset
539 int h = (mpi->h + 15) >> 4;
915be5c7a30c Make sure that a qstride of 0 (intentional or not) does not completely break
reimar
parents: 29310
diff changeset
540 if (!w) {
915be5c7a30c Make sure that a qstride of 0 (intentional or not) does not completely break
reimar
parents: 29310
diff changeset
541 w = (mpi->w + 15) >> 4;
915be5c7a30c Make sure that a qstride of 0 (intentional or not) does not completely break
reimar
parents: 29310
diff changeset
542 h = 1;
915be5c7a30c Make sure that a qstride of 0 (intentional or not) does not completely break
reimar
parents: 29310
diff changeset
543 }
915be5c7a30c Make sure that a qstride of 0 (intentional or not) does not completely break
reimar
parents: 29310
diff changeset
544 if(!vf->priv->non_b_qp)
915be5c7a30c Make sure that a qstride of 0 (intentional or not) does not completely break
reimar
parents: 29310
diff changeset
545 vf->priv->non_b_qp= malloc(w*h);
915be5c7a30c Make sure that a qstride of 0 (intentional or not) does not completely break
reimar
parents: 29310
diff changeset
546 fast_memcpy(vf->priv->non_b_qp, mpi->qscale, w*h);
17133
a2b24e0d7772 prevent flicker on b-frames, trivial port from vf_spp
henry
parents: 17012
diff changeset
547 }
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
548 if(vf->priv->log2_count || !(mpi->flags&MP_IMGFLAG_DIRECT)){
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
549 char *qp_tab= vf->priv->non_b_qp;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
550 if(vf->priv->bframes || !qp_tab)
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
551 qp_tab= mpi->qscale;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
552
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
553 if(qp_tab || vf->priv->qp){
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
554 filter(vf->priv, dmpi->planes[0], mpi->planes[0], dmpi->stride[0], mpi->stride[0],
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
555 mpi->w, mpi->h, qp_tab, mpi->qstride, 1);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
556 filter(vf->priv, dmpi->planes[1], mpi->planes[1], dmpi->stride[1], mpi->stride[1],
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
557 mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, qp_tab, mpi->qstride, 0);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
558 filter(vf->priv, dmpi->planes[2], mpi->planes[2], dmpi->stride[2], mpi->stride[2],
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
559 mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, qp_tab, mpi->qstride, 0);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
560 }else{
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
561 memcpy_pic(dmpi->planes[0], mpi->planes[0], mpi->w, mpi->h, dmpi->stride[0], mpi->stride[0]);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
562 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]);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
563 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]);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
564 }
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
28290
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_...
reimar
parents: 27754
diff changeset
567 #if HAVE_MMX
27754
08d18fe9da52 Change all occurrences of asm and __asm to __asm__, same as was done for FFmpeg.
diego
parents: 26727
diff changeset
568 if(gCpuCaps.hasMMX) __asm__ volatile ("emms\n\t");
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
569 #endif
28290
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_...
reimar
parents: 27754
diff changeset
570 #if HAVE_MMX2
27754
08d18fe9da52 Change all occurrences of asm and __asm to __asm__, same as was done for FFmpeg.
diego
parents: 26727
diff changeset
571 if(gCpuCaps.hasMMX2) __asm__ volatile ("sfence\n\t");
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
572 #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
573 return vf_next_put_image(vf,dmpi, pts);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
574 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
575
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
576 static void uninit(struct vf_instance *vf)
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
577 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
578 if(!vf->priv) return;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
579
32537
8fa2f43cb760 Remove most of the NULL pointer check before free all over the code
cboesch
parents: 31959
diff changeset
580 av_free(vf->priv->temp);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
581 vf->priv->temp= NULL;
32537
8fa2f43cb760 Remove most of the NULL pointer check before free all over the code
cboesch
parents: 31959
diff changeset
582 av_free(vf->priv->src);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
583 vf->priv->src= NULL;
32537
8fa2f43cb760 Remove most of the NULL pointer check before free all over the code
cboesch
parents: 31959
diff changeset
584 //free(vf->priv->avctx);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
585 //vf->priv->avctx= NULL;
32537
8fa2f43cb760 Remove most of the NULL pointer check before free all over the code
cboesch
parents: 31959
diff changeset
586 free(vf->priv->non_b_qp);
17133
a2b24e0d7772 prevent flicker on b-frames, trivial port from vf_spp
henry
parents: 17012
diff changeset
587 vf->priv->non_b_qp= NULL;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
588
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
589 av_free(vf->priv);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
590 vf->priv=NULL;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
591 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
592
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
593 //===========================================================================//
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
594
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
595 static int query_format(struct vf_instance *vf, unsigned int fmt)
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
596 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
597 switch(fmt){
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
598 case IMGFMT_YVU9:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
599 case IMGFMT_IF09:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
600 case IMGFMT_YV12:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
601 case IMGFMT_I420:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
602 case IMGFMT_IYUV:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
603 case IMGFMT_CLPL:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
604 case IMGFMT_Y800:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
605 case IMGFMT_Y8:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
606 case IMGFMT_444P:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
607 case IMGFMT_422P:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
608 case IMGFMT_411P:
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
609 return vf_next_query_format(vf,fmt);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
610 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
611 return 0;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
612 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
613
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
614 static int control(struct vf_instance *vf, int request, void* data)
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
615 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
616 switch(request){
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
617 case VFCTRL_QUERY_MAX_PP_LEVEL:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
618 return 5;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
619 case VFCTRL_SET_PP_LEVEL:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
620 vf->priv->log2_count= *((unsigned int*)data);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
621 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
622 return CONTROL_TRUE;
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 return vf_next_control(vf,request,data);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
625 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
626
30638
a7b908875c14 Rename open() vf initialization function to vf_open().
diego
parents: 30412
diff changeset
627 static int vf_open(vf_instance_t *vf, char *args)
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
628 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
629 int i=0, bias;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
630 int custom_threshold_m[64];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
631 int log2c=-1;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
632
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
633 vf->config=config;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
634 vf->put_image=put_image;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
635 vf->get_image=get_image;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
636 vf->query_format=query_format;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
637 vf->uninit=uninit;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
638 vf->control= control;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
639 vf->priv=av_mallocz(sizeof(struct vf_priv_s));//assumes align 16 !
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
640
31959
f957f330aa6d Introduce init_avcodec function to avoid duplicated FFmpeg initializations.
diego
parents: 30642
diff changeset
641 init_avcodec();
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
642
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
643 //vf->priv->avctx= avcodec_alloc_context();
34728
a28c4ba97171 Replace use of deprecated dsputil_init() by ff_dsputil_init().
diego
parents: 34198
diff changeset
644 //ff_dsputil_init(&vf->priv->dsp, vf->priv->avctx);
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
645
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
646 vf->priv->log2_count= 4;
17225
ec9888363742 reverse the H264 hack
henry
parents: 17223
diff changeset
647 vf->priv->bframes = 0;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
648
17225
ec9888363742 reverse the H264 hack
henry
parents: 17223
diff changeset
649 if (args) sscanf(args, "%d:%d:%d:%d", &log2c, &vf->priv->qp, &i, &vf->priv->bframes);
ec9888363742 reverse the H264 hack
henry
parents: 17223
diff changeset
650
ec9888363742 reverse the H264 hack
henry
parents: 17223
diff changeset
651 if( log2c >=4 && log2c <=5 )
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
652 vf->priv->log2_count = log2c;
15651
6a0494e09435 sanity checks for options; treat quality > 5 as 5, not 4
henry
parents: 15634
diff changeset
653 else if( log2c >= 6 )
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
654 vf->priv->log2_count = 5;
15631
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 if(vf->priv->qp < 0)
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
657 vf->priv->qp = 0;
15651
6a0494e09435 sanity checks for options; treat quality > 5 as 5, not 4
henry
parents: 15634
diff changeset
658
6a0494e09435 sanity checks for options; treat quality > 5 as 5, not 4
henry
parents: 15634
diff changeset
659 if (i < -15) i = -15;
6a0494e09435 sanity checks for options; treat quality > 5 as 5, not 4
henry
parents: 15634
diff changeset
660 if (i > 32) i = 32;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
661
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
662 bias= (1<<4)+i; //regulable
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
663 vf->priv->prev_q=0;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
664 //
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
665 for(i=0;i<64;i++) //FIXME: tune custom_threshold[] and remove this !
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
666 custom_threshold_m[i]=(int)(custom_threshold[i]*(bias/71.)+ 0.5);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
667 for(i=0;i<8;i++){
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
668 vf->priv->threshold_mtx_noq[2*i]=(uint64_t)custom_threshold_m[i*8+2]
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
669 |(((uint64_t)custom_threshold_m[i*8+6])<<16)
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
670 |(((uint64_t)custom_threshold_m[i*8+0])<<32)
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
671 |(((uint64_t)custom_threshold_m[i*8+4])<<48);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
672 vf->priv->threshold_mtx_noq[2*i+1]=(uint64_t)custom_threshold_m[i*8+5]
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
673 |(((uint64_t)custom_threshold_m[i*8+3])<<16)
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
674 |(((uint64_t)custom_threshold_m[i*8+1])<<32)
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
675 |(((uint64_t)custom_threshold_m[i*8+7])<<48);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
676 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
677
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
678 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
679
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
680 return 1;
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
25221
00fff9a3b735 Make all vf_info_t structs const
reimar
parents: 24976
diff changeset
683 const vf_info_t vf_info_fspp = {
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
684 "fast simple postprocess",
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
685 "fspp",
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
686 "Michael Niedermayer, Nikolaj Poroshin",
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
687 "",
30638
a7b908875c14 Rename open() vf initialization function to vf_open().
diego
parents: 30412
diff changeset
688 vf_open,
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
689 NULL
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
690 };
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
691
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
692 //====================================================================
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
693 //Specific spp's dct, idct and threshold functions
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
694 //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
695
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
696 //#define MANGLE(a) #a
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
697
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
698 //typedef int16_t DCTELEM; //! only int16_t
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
699
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
700 #define DCTSIZE 8
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
701 #define DCTSIZE_S "8"
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 #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
704 #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
705 #define FIX64(x,s) C64(FIX(x,s))
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
706
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
707 #define MULTIPLY16H(x,k) (((x)*(k))>>16)
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
708 #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
709 #define DESCALE(x,n) (((x) + (1 << ((n)-1))) >> n)
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
710
28290
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_...
reimar
parents: 27754
diff changeset
711 #if HAVE_MMX
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
712
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
713 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_382683433)=FIX64(0.382683433, 14);
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
714 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_541196100)=FIX64(0.541196100, 14);
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
715 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_707106781)=FIX64(0.707106781, 14);
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
716 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_306562965)=FIX64(1.306562965, 14);
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
717
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
718 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_414213562_A)=FIX64(1.414213562, 14);
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
719
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
720 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_847759065)=FIX64(1.847759065, 13);
25901
c2210e68a2a9 Simplify: use DECLARE_ASM_CONST
reimar
parents: 25568
diff changeset
721 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_2_613125930)=FIX64(-2.613125930, 13); //-
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
722 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_414213562)=FIX64(1.414213562, 13);
25901
c2210e68a2a9 Simplify: use DECLARE_ASM_CONST
reimar
parents: 25568
diff changeset
723 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
724 //for t3,t5,t7 == 0 shortcut
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
725 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_847759065)=FIX64(0.847759065, 14);
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
726 DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_566454497)=FIX64(0.566454497, 14);
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
727 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
728
25901
c2210e68a2a9 Simplify: use DECLARE_ASM_CONST
reimar
parents: 25568
diff changeset
729 DECLARE_ASM_CONST(8, uint64_t, MM_DESCALE_RND)=C64(4);
c2210e68a2a9 Simplify: use DECLARE_ASM_CONST
reimar
parents: 25568
diff changeset
730 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
731
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
732 #else /* !HAVE_MMX */
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
733
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
734 typedef int32_t int_simd16_t;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
735 static const int16_t FIX_0_382683433=FIX(0.382683433, 14);
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
736 static const int16_t FIX_0_541196100=FIX(0.541196100, 14);
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
737 static const int16_t FIX_0_707106781=FIX(0.707106781, 14);
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
738 static const int16_t FIX_1_306562965=FIX(1.306562965, 14);
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
739 static const int16_t FIX_1_414213562_A=FIX(1.414213562, 14);
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
740 static const int16_t FIX_1_847759065=FIX(1.847759065, 13);
25902
15ab840747e2 mark constants as such
reimar
parents: 25901
diff changeset
741 static const int16_t FIX_2_613125930=FIX(-2.613125930, 13); //-
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
742 static const int16_t FIX_1_414213562=FIX(1.414213562, 13);
25902
15ab840747e2 mark constants as such
reimar
parents: 25901
diff changeset
743 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
744
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
745 #endif
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
746
28290
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_...
reimar
parents: 27754
diff changeset
747 #if !HAVE_MMX
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
748
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
749 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
750 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
751 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
752 int_simd16_t tmp10, tmp11, tmp12, tmp13;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
753 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
754 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
755
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
756 DCTELEM* dataptr;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
757 DCTELEM* wsptr;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
758 int16_t *threshold;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
759 int ctr;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
760
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
761 dataptr = data;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
762 wsptr = output;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
763
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
764 for (; cnt > 0; cnt-=2) { //start positions
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
765 threshold=(int16_t*)thr_adr;//threshold_mtx
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
766 for (ctr = DCTSIZE; ctr > 0; ctr--) {
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
767 // Process columns from input, add to output.
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
768 tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
769 tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
770
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
771 tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
772 tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
773
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
774 tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
775 tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
776
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
777 tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
778 tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
779
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
780 // Even part of FDCT
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
781
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
782 tmp10 = tmp0 + tmp3;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
783 tmp13 = tmp0 - tmp3;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
784 tmp11 = tmp1 + tmp2;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
785 tmp12 = tmp1 - tmp2;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
786
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
787 d0 = tmp10 + tmp11;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
788 d4 = tmp10 - tmp11;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
789
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
790 z1 = MULTIPLY16H((tmp12 + tmp13) <<2, FIX_0_707106781);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
791 d2 = tmp13 + z1;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
792 d6 = tmp13 - z1;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
793
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
794 // Even part of IDCT
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
795
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
796 THRESHOLD(tmp0, d0, threshold[0*8]);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
797 THRESHOLD(tmp1, d2, threshold[2*8]);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
798 THRESHOLD(tmp2, d4, threshold[4*8]);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
799 THRESHOLD(tmp3, d6, threshold[6*8]);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
800 tmp0+=2;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
801 tmp10 = (tmp0 + tmp2)>>2;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
802 tmp11 = (tmp0 - tmp2)>>2;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
803
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
804 tmp13 = (tmp1 + tmp3)>>2; //+2 ! (psnr decides)
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
805 tmp12 = MULTIPLY16H((tmp1 - tmp3), FIX_1_414213562_A) - tmp13; //<<2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
806
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
807 tmp0 = tmp10 + tmp13; //->temps
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
808 tmp3 = tmp10 - tmp13; //->temps
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
809 tmp1 = tmp11 + tmp12; //->temps
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
810 tmp2 = tmp11 - tmp12; //->temps
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
811
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
812 // Odd part of FDCT
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
813
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
814 tmp10 = tmp4 + tmp5;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
815 tmp11 = tmp5 + tmp6;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
816 tmp12 = tmp6 + tmp7;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
817
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
818 z5 = MULTIPLY16H((tmp10 - tmp12)<<2, FIX_0_382683433);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
819 z2 = MULTIPLY16H(tmp10 <<2, FIX_0_541196100) + z5;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
820 z4 = MULTIPLY16H(tmp12 <<2, FIX_1_306562965) + z5;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
821 z3 = MULTIPLY16H(tmp11 <<2, FIX_0_707106781);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
822
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
823 z11 = tmp7 + z3;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
824 z13 = tmp7 - z3;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
825
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
826 d5 = z13 + z2;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
827 d3 = z13 - z2;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
828 d1 = z11 + z4;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
829 d7 = z11 - z4;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
830
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
831 // Odd part of IDCT
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
832
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
833 THRESHOLD(tmp4, d1, threshold[1*8]);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
834 THRESHOLD(tmp5, d3, threshold[3*8]);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
835 THRESHOLD(tmp6, d5, threshold[5*8]);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
836 THRESHOLD(tmp7, d7, threshold[7*8]);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
837
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
838 //Simd version uses here a shortcut for the tmp5,tmp6,tmp7 == 0
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
839 z13 = tmp6 + tmp5;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
840 z10 = (tmp6 - tmp5)<<1;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
841 z11 = tmp4 + tmp7;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
842 z12 = (tmp4 - tmp7)<<1;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
843
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
844 tmp7 = (z11 + z13)>>2; //+2 !
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
845 tmp11 = MULTIPLY16H((z11 - z13)<<1, FIX_1_414213562);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
846 z5 = MULTIPLY16H(z10 + z12, FIX_1_847759065);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
847 tmp10 = MULTIPLY16H(z12, FIX_1_082392200) - z5;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
848 tmp12 = MULTIPLY16H(z10, FIX_2_613125930) + z5; // - !!
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
849
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
850 tmp6 = tmp12 - tmp7;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
851 tmp5 = tmp11 - tmp6;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
852 tmp4 = tmp10 + tmp5;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
853
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
854 wsptr[DCTSIZE*0]+= (tmp0 + tmp7);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
855 wsptr[DCTSIZE*1]+= (tmp1 + tmp6);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
856 wsptr[DCTSIZE*2]+= (tmp2 + tmp5);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
857 wsptr[DCTSIZE*3]+= (tmp3 - tmp4);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
858 wsptr[DCTSIZE*4]+= (tmp3 + tmp4);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
859 wsptr[DCTSIZE*5]+= (tmp2 - tmp5);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
860 wsptr[DCTSIZE*6]= (tmp1 - tmp6);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
861 wsptr[DCTSIZE*7]= (tmp0 - tmp7);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
862 //
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
863 dataptr++; //next column
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
864 wsptr++;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
865 threshold++;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
866 }
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
867 dataptr+=8; //skip each second start pos
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
868 wsptr +=8;
15631
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 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
871
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
872 #else /* HAVE_MMX */
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
873
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
874 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
875 {
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
876 uint64_t __attribute__((aligned(8))) temps[4];
27754
08d18fe9da52 Change all occurrences of asm and __asm to __asm__, same as was done for FFmpeg.
diego
parents: 26727
diff changeset
877 __asm__ volatile(
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
878 ASMALIGN(4)
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
879 "1: \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
880 "movq "DCTSIZE_S"*0*2(%%"REG_S"), %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
881 //
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
882 "movq "DCTSIZE_S"*3*2(%%"REG_S"), %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
883 "movq %%mm1, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
884
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
885 "paddw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm1 \n\t" //t0
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
886 "movq %%mm7, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
887
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
888 "paddw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm7 \n\t" //t3
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
889 "movq %%mm1, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
890
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
891 "movq "DCTSIZE_S"*1*2(%%"REG_S"), %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
892 "psubw %%mm7, %%mm1 \n\t" //t13
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
893
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
894 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
895 "movq %%mm6, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
896
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
897 "paddw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm6 \n\t" //t1
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
898 "paddw %%mm7, %%mm5 \n\t" //t10
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
899
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
900 "paddw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
901 "movq %%mm6, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
902
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
903 "paddw %%mm2, %%mm6 \n\t" //t11
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
904 "psubw %%mm2, %%mm7 \n\t" //t12
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
905
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
906 "movq %%mm5, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
907 "paddw %%mm6, %%mm5 \n\t" //d0
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
908 // i0 t13 t12 i3 i1 d0 - d4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
909 "psubw %%mm6, %%mm2 \n\t" //d4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
910 "paddw %%mm1, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
911
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
912 "movq 4*16(%%"REG_d"), %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
913 "psllw $2, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
914
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
915 "psubw 0*16(%%"REG_d"), %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
916 "psubw %%mm6, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
917
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
918 "paddusw 0*16(%%"REG_d"), %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
919 "paddusw %%mm6, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
920
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
921 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
922 //
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
923 "paddw 0*16(%%"REG_d"), %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
924 "paddw %%mm6, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
925
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
926 "psubusw 0*16(%%"REG_d"), %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
927 "psubusw %%mm6, %%mm2 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
928
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
929 //This func is totally compute-bound, operates at huge speed. So, DC shortcut
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
930 // at this place isn't worthwhile due to BTB miss penalty (checked on Pent. 3).
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
931 //However, typical numbers: nondc - 29%%, dc - 46%%, zero - 25%%. All <> 0 case is very rare.
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
932 "paddw "MANGLE(MM_2)", %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
933 "movq %%mm2, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
934
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
935 "paddw %%mm5, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
936 "psubw %%mm6, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
937
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
938 "movq %%mm1, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
939 "paddw %%mm7, %%mm1 \n\t" //d2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
940
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
941 "psubw 2*16(%%"REG_d"), %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
942 "psubw %%mm7, %%mm6 \n\t" //d6
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
943
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
944 "movq 6*16(%%"REG_d"), %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
945 "psraw $2, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
946
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
947 "paddusw 2*16(%%"REG_d"), %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
948 "psubw %%mm7, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
949 // t7 d2 /t11 t4 t6 - d6 /t10
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
950
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
951 "paddw 2*16(%%"REG_d"), %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
952 "paddusw %%mm7, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
953
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
954 "psubusw 2*16(%%"REG_d"), %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
955 "paddw %%mm7, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
956
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
957 "psubw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
958 "psubusw %%mm7, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
959
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
960 //movq [edi+"DCTSIZE_S"*2*2], mm1
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
961 //movq [edi+"DCTSIZE_S"*6*2], mm6
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
962 "movq %%mm1, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
963 "psraw $2, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
964
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
965 "psubw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
966 "psubw %%mm6, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
967
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
968 "psubw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
969 "paddw %%mm7, %%mm6 \n\t" //'t13
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
970
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
971 "psraw $2, %%mm6 \n\t" //paddw mm6, MM_2 !! ---
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
972 "movq %%mm2, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
973
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
974 "pmulhw "MANGLE(MM_FIX_1_414213562_A)", %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
975 "paddw %%mm6, %%mm2 \n\t" //'t0
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
976
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
977 "movq %%mm2, 0*8+%3 \n\t" //!
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
978 "psubw %%mm6, %%mm7 \n\t" //'t3
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
979
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
980 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
981 "psubw %%mm6, %%mm1 \n\t" //'t12
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
982
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
983 "psubw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t5
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
984 "movq %%mm5, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
985
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
986 "movq %%mm7, 3*8+%3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
987 "paddw %%mm2, %%mm3 \n\t" //t10
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
988
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
989 "paddw %%mm4, %%mm2 \n\t" //t11
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
990 "paddw %%mm0, %%mm4 \n\t" //t12
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
991
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
992 "movq %%mm3, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
993 "psubw %%mm4, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
994
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
995 "psllw $2, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
996 "psllw $2, %%mm7 \n\t" //opt for P6
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
997
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
998 "pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
999 "psllw $2, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1000
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1001 "pmulhw "MANGLE(MM_FIX_0_541196100)", %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1002 "psllw $2, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1003
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1004 "pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1005 "paddw %%mm1, %%mm5 \n\t" //'t1
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1006
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1007 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1008 "psubw %%mm1, %%mm6 \n\t" //'t2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1009 // t7 't12 't11 t4 t6 - 't13 't10 ---
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1010
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1011 "paddw %%mm3, %%mm7 \n\t" //z2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1012
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1013 "movq %%mm5, 1*8+%3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1014 "paddw %%mm3, %%mm4 \n\t" //z4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1015
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1016 "movq 3*16(%%"REG_d"), %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1017 "movq %%mm0, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1018
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1019 "movq %%mm6, 2*8+%3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1020 "psubw %%mm2, %%mm1 \n\t" //z13
15631
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 //===
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1023 "paddw %%mm2, %%mm0 \n\t" //z11
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1024 "movq %%mm1, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1025
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1026 "movq 5*16(%%"REG_d"), %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1027 "psubw %%mm7, %%mm1 \n\t" //d3
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1028
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1029 "paddw %%mm7, %%mm5 \n\t" //d5
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1030 "psubw %%mm3, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1031
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1032 "movq 1*16(%%"REG_d"), %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1033 "psubw %%mm2, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1034
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1035 "movq %%mm0, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1036 "paddw %%mm4, %%mm0 \n\t" //d1
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1037
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1038 "paddusw %%mm3, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1039 "psubw %%mm4, %%mm6 \n\t" //d7
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1040
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1041 // d1 d3 - - - d5 d7 -
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1042 "movq 7*16(%%"REG_d"), %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1043 "psubw %%mm7, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1044
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1045 "psubw %%mm4, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1046 "paddusw %%mm2, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1047
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1048 "paddusw %%mm4, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1049 "paddw %%mm3, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1050
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1051 "paddw %%mm2, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1052 "paddw %%mm4, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1053
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1054 "psubusw %%mm3, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1055 "psubusw %%mm2, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1056
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1057 "psubusw %%mm4, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1058 "movq %%mm1, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1059
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1060 "por %%mm5, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1061 "paddusw %%mm7, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1062
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1063 "por %%mm6, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1064 "paddw %%mm7, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1065
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1066 "packssdw %%mm4, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1067 "psubusw %%mm7, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1068
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1069 "movd %%mm4, %%"REG_a" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1070 "or %%"REG_a", %%"REG_a" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1071 "jnz 2f \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1072 //movq [edi+"DCTSIZE_S"*3*2], mm1
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1073 //movq [edi+"DCTSIZE_S"*5*2], mm5
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1074 //movq [edi+"DCTSIZE_S"*1*2], mm0
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1075 //movq [edi+"DCTSIZE_S"*7*2], mm6
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1076 // t4 t5 - - - t6 t7 -
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1077 //--- t4 (mm0) may be <>0; mm1, mm5, mm6 == 0
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1078 //Typical numbers: nondc - 19%%, dc - 26%%, zero - 55%%. zero case alone isn't worthwhile
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1079 "movq 0*8+%3, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1080 "movq %%mm0, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1081
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1082 "pmulhw "MANGLE(MM_FIX_0_847759065)", %%mm0 \n\t" //tmp6
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1083 "movq %%mm1, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1084
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1085 "movq "DCTSIZE_S"*0*2(%%"REG_D"), %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1086 "movq %%mm2, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1087
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1088 "pmulhw "MANGLE(MM_FIX_0_566454497)", %%mm1 \n\t" //tmp5
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1089 "paddw %%mm4, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1090
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1091 "movq 1*8+%3, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1092 //paddw mm3, MM_2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1093 "psraw $2, %%mm3 \n\t" //tmp7
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1094
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1095 "pmulhw "MANGLE(MM_FIX_0_198912367)", %%mm2 \n\t" //-tmp4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1096 "psubw %%mm3, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1097
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1098 "movq "DCTSIZE_S"*1*2(%%"REG_D"), %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1099 "paddw %%mm3, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1100
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1101 "movq %%mm4, "DCTSIZE_S"*7*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1102 "paddw %%mm6, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1103
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1104 "movq 2*8+%3, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1105 "psubw %%mm0, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1106
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1107 "movq "DCTSIZE_S"*2*2(%%"REG_D"), %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1108 "paddw %%mm0, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1109
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1110 "movq %%mm5, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1111 "paddw %%mm3, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1112
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1113 "movq %%mm6, "DCTSIZE_S"*6*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1114 "psubw %%mm1, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1115
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1116 "movq "DCTSIZE_S"*5*2(%%"REG_D"), %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1117 "paddw %%mm1, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1118
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1119 "movq "DCTSIZE_S"*3*2(%%"REG_D"), %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1120 "paddw %%mm3, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1121
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1122 "movq 3*8+%3, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1123 "add $8, %%"REG_S" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1124
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1125 "movq %%mm7, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1126 "paddw %%mm0, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1127
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1128 "movq %%mm4, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1129 "psubw %%mm2, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1130
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1131 "movq "DCTSIZE_S"*4*2(%%"REG_D"), %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1132 "paddw %%mm2, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1133
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1134 "movq %%mm5, "DCTSIZE_S"*5*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1135 "paddw %%mm0, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1136
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1137 "movq %%mm6, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1138
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1139 "movq %%mm7, "DCTSIZE_S"*4*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1140 "add $8, %%"REG_D" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1141 "jmp 4f \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1142
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1143 "2: \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1144 //--- non DC2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1145 //psraw mm1, 2 w/o it -> offset. thr1, thr1, thr1 (actually thr1, thr1, thr1-1)
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1146 //psraw mm5, 2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1147 //psraw mm0, 2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1148 //psraw mm6, 2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1149 "movq %%mm5, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1150 "psubw %%mm1, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1151
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1152 "psllw $1, %%mm5 \n\t" //'z10
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1153 "paddw %%mm1, %%mm3 \n\t" //'z13
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1154
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1155 "movq %%mm0, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1156 "psubw %%mm6, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1157
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1158 "movq %%mm5, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1159 "psllw $1, %%mm0 \n\t" //'z12
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1160
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1161 "pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm1 \n\t" //-
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1162 "paddw %%mm0, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1163
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1164 "pmulhw "MANGLE(MM_FIX_1_847759065)", %%mm5 \n\t" //'z5
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1165 "paddw %%mm6, %%mm2 \n\t" //'z11
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1166
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1167 "pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1168 "movq %%mm2, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1169
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1170 //---
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1171 "movq 0*8+%3, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1172 "psubw %%mm3, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1173
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1174 "psllw $1, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1175 "paddw %%mm3, %%mm7 \n\t" //'t7
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1176
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1177 "pmulhw "MANGLE(MM_FIX_1_414213562)", %%mm2 \n\t" //'t11
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1178 "movq %%mm4, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1179 //paddw mm7, MM_2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1180 "psraw $2, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1181
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1182 "paddw "DCTSIZE_S"*0*2(%%"REG_D"), %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1183 "psubw %%mm7, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1184
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1185 "movq 1*8+%3, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1186 "paddw %%mm7, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1187
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1188 "movq %%mm6, "DCTSIZE_S"*7*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1189 "paddw %%mm5, %%mm1 \n\t" //'t12
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1190
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1191 "movq %%mm4, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1192 "psubw %%mm7, %%mm1 \n\t" //'t6
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1193
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1194 "movq 2*8+%3, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1195 "psubw %%mm5, %%mm0 \n\t" //'t10
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1196
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1197 "movq 3*8+%3, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1198 "movq %%mm3, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1199
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1200 "paddw "DCTSIZE_S"*1*2(%%"REG_D"), %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1201 "psubw %%mm1, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1202
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1203 "psubw %%mm1, %%mm2 \n\t" //'t5
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1204 "paddw %%mm1, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1205
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1206 "movq %%mm5, "DCTSIZE_S"*6*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1207 "movq %%mm7, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1208
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1209 "paddw "DCTSIZE_S"*2*2(%%"REG_D"), %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1210 "psubw %%mm2, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1211
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1212 "paddw "DCTSIZE_S"*5*2(%%"REG_D"), %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1213 "paddw %%mm2, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1214
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1215 "movq %%mm3, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1216 "paddw %%mm2, %%mm0 \n\t" //'t4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1217
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1218 // 't4 't6 't5 - - - - 't7
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1219 "movq %%mm7, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1220 "movq %%mm6, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1221
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1222 "paddw "DCTSIZE_S"*4*2(%%"REG_D"), %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1223 "psubw %%mm0, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1224
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1225 "paddw "DCTSIZE_S"*3*2(%%"REG_D"), %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1226 "paddw %%mm0, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1227
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1228 "movq %%mm4, "DCTSIZE_S"*5*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1229 "add $8, %%"REG_S" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1230
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1231 "movq %%mm6, "DCTSIZE_S"*4*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1232
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1233 "movq %%mm1, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1234 "add $8, %%"REG_D" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1235
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1236 "4: \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1237 //=part 2 (the same)===========================================================
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1238 "movq "DCTSIZE_S"*0*2(%%"REG_S"), %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1239 //
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1240 "movq "DCTSIZE_S"*3*2(%%"REG_S"), %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1241 "movq %%mm1, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1242
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1243 "paddw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm1 \n\t" //t0
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1244 "movq %%mm7, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1245
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1246 "paddw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm7 \n\t" //t3
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1247 "movq %%mm1, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1248
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1249 "movq "DCTSIZE_S"*1*2(%%"REG_S"), %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1250 "psubw %%mm7, %%mm1 \n\t" //t13
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1251
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1252 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1253 "movq %%mm6, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1254
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1255 "paddw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm6 \n\t" //t1
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1256 "paddw %%mm7, %%mm5 \n\t" //t10
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1257
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1258 "paddw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1259 "movq %%mm6, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1260
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1261 "paddw %%mm2, %%mm6 \n\t" //t11
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1262 "psubw %%mm2, %%mm7 \n\t" //t12
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1263
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1264 "movq %%mm5, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1265 "paddw %%mm6, %%mm5 \n\t" //d0
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1266 // i0 t13 t12 i3 i1 d0 - d4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1267 "psubw %%mm6, %%mm2 \n\t" //d4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1268 "paddw %%mm1, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1269
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1270 "movq 1*8+4*16(%%"REG_d"), %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1271 "psllw $2, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1272
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1273 "psubw 1*8+0*16(%%"REG_d"), %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1274 "psubw %%mm6, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1275
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1276 "paddusw 1*8+0*16(%%"REG_d"), %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1277 "paddusw %%mm6, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1278
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1279 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1280 //
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1281 "paddw 1*8+0*16(%%"REG_d"), %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1282 "paddw %%mm6, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1283
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1284 "psubusw 1*8+0*16(%%"REG_d"), %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1285 "psubusw %%mm6, %%mm2 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1286
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1287 //This func is totally compute-bound, operates at huge speed. So, DC shortcut
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1288 // at this place isn't worthwhile due to BTB miss penalty (checked on Pent. 3).
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1289 //However, typical numbers: nondc - 29%%, dc - 46%%, zero - 25%%. All <> 0 case is very rare.
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1290 "paddw "MANGLE(MM_2)", %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1291 "movq %%mm2, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1292
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1293 "paddw %%mm5, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1294 "psubw %%mm6, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1295
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1296 "movq %%mm1, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1297 "paddw %%mm7, %%mm1 \n\t" //d2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1298
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1299 "psubw 1*8+2*16(%%"REG_d"), %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1300 "psubw %%mm7, %%mm6 \n\t" //d6
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1301
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1302 "movq 1*8+6*16(%%"REG_d"), %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1303 "psraw $2, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1304
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1305 "paddusw 1*8+2*16(%%"REG_d"), %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1306 "psubw %%mm7, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1307 // t7 d2 /t11 t4 t6 - d6 /t10
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1308
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1309 "paddw 1*8+2*16(%%"REG_d"), %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1310 "paddusw %%mm7, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1311
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1312 "psubusw 1*8+2*16(%%"REG_d"), %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1313 "paddw %%mm7, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1314
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1315 "psubw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1316 "psubusw %%mm7, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1317
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1318 //movq [edi+"DCTSIZE_S"*2*2], mm1
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1319 //movq [edi+"DCTSIZE_S"*6*2], mm6
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1320 "movq %%mm1, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1321 "psraw $2, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1322
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1323 "psubw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1324 "psubw %%mm6, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1325
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1326 "psubw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1327 "paddw %%mm7, %%mm6 \n\t" //'t13
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1328
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1329 "psraw $2, %%mm6 \n\t" //paddw mm6, MM_2 !! ---
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1330 "movq %%mm2, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1331
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1332 "pmulhw "MANGLE(MM_FIX_1_414213562_A)", %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1333 "paddw %%mm6, %%mm2 \n\t" //'t0
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1334
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1335 "movq %%mm2, 0*8+%3 \n\t" //!
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1336 "psubw %%mm6, %%mm7 \n\t" //'t3
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1337
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1338 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1339 "psubw %%mm6, %%mm1 \n\t" //'t12
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1340
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1341 "psubw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t5
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1342 "movq %%mm5, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1343
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1344 "movq %%mm7, 3*8+%3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1345 "paddw %%mm2, %%mm3 \n\t" //t10
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1346
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1347 "paddw %%mm4, %%mm2 \n\t" //t11
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1348 "paddw %%mm0, %%mm4 \n\t" //t12
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1349
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1350 "movq %%mm3, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1351 "psubw %%mm4, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1352
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1353 "psllw $2, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1354 "psllw $2, %%mm7 \n\t" //opt for P6
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1355
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1356 "pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1357 "psllw $2, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1358
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1359 "pmulhw "MANGLE(MM_FIX_0_541196100)", %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1360 "psllw $2, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1361
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1362 "pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1363 "paddw %%mm1, %%mm5 \n\t" //'t1
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1364
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1365 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1366 "psubw %%mm1, %%mm6 \n\t" //'t2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1367 // t7 't12 't11 t4 t6 - 't13 't10 ---
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1368
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1369 "paddw %%mm3, %%mm7 \n\t" //z2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1370
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1371 "movq %%mm5, 1*8+%3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1372 "paddw %%mm3, %%mm4 \n\t" //z4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1373
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1374 "movq 1*8+3*16(%%"REG_d"), %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1375 "movq %%mm0, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1376
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1377 "movq %%mm6, 2*8+%3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1378 "psubw %%mm2, %%mm1 \n\t" //z13
15631
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 //===
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1381 "paddw %%mm2, %%mm0 \n\t" //z11
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1382 "movq %%mm1, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1383
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1384 "movq 1*8+5*16(%%"REG_d"), %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1385 "psubw %%mm7, %%mm1 \n\t" //d3
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1386
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1387 "paddw %%mm7, %%mm5 \n\t" //d5
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1388 "psubw %%mm3, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1389
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1390 "movq 1*8+1*16(%%"REG_d"), %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1391 "psubw %%mm2, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1392
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1393 "movq %%mm0, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1394 "paddw %%mm4, %%mm0 \n\t" //d1
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1395
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1396 "paddusw %%mm3, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1397 "psubw %%mm4, %%mm6 \n\t" //d7
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1398
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1399 // d1 d3 - - - d5 d7 -
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1400 "movq 1*8+7*16(%%"REG_d"), %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1401 "psubw %%mm7, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1402
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1403 "psubw %%mm4, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1404 "paddusw %%mm2, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1405
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1406 "paddusw %%mm4, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1407 "paddw %%mm3, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1408
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1409 "paddw %%mm2, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1410 "paddw %%mm4, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1411
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1412 "psubusw %%mm3, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1413 "psubusw %%mm2, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1414
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1415 "psubusw %%mm4, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1416 "movq %%mm1, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1417
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1418 "por %%mm5, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1419 "paddusw %%mm7, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1420
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1421 "por %%mm6, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1422 "paddw %%mm7, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1423
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1424 "packssdw %%mm4, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1425 "psubusw %%mm7, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1426
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1427 "movd %%mm4, %%"REG_a" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1428 "or %%"REG_a", %%"REG_a" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1429 "jnz 3f \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1430 //movq [edi+"DCTSIZE_S"*3*2], mm1
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1431 //movq [edi+"DCTSIZE_S"*5*2], mm5
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1432 //movq [edi+"DCTSIZE_S"*1*2], mm0
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1433 //movq [edi+"DCTSIZE_S"*7*2], mm6
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1434 // t4 t5 - - - t6 t7 -
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1435 //--- t4 (mm0) may be <>0; mm1, mm5, mm6 == 0
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1436 //Typical numbers: nondc - 19%%, dc - 26%%, zero - 55%%. zero case alone isn't worthwhile
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1437 "movq 0*8+%3, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1438 "movq %%mm0, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1439
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1440 "pmulhw "MANGLE(MM_FIX_0_847759065)", %%mm0 \n\t" //tmp6
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1441 "movq %%mm1, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1442
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1443 "movq "DCTSIZE_S"*0*2(%%"REG_D"), %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1444 "movq %%mm2, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1445
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1446 "pmulhw "MANGLE(MM_FIX_0_566454497)", %%mm1 \n\t" //tmp5
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1447 "paddw %%mm4, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1448
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1449 "movq 1*8+%3, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1450 //paddw mm3, MM_2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1451 "psraw $2, %%mm3 \n\t" //tmp7
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1452
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1453 "pmulhw "MANGLE(MM_FIX_0_198912367)", %%mm2 \n\t" //-tmp4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1454 "psubw %%mm3, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1455
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1456 "movq "DCTSIZE_S"*1*2(%%"REG_D"), %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1457 "paddw %%mm3, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1458
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1459 "movq %%mm4, "DCTSIZE_S"*7*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1460 "paddw %%mm6, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1461
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1462 "movq 2*8+%3, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1463 "psubw %%mm0, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1464
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1465 "movq "DCTSIZE_S"*2*2(%%"REG_D"), %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1466 "paddw %%mm0, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1467
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1468 "movq %%mm5, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1469 "paddw %%mm3, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1470
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1471 "movq %%mm6, "DCTSIZE_S"*6*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1472 "psubw %%mm1, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1473
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1474 "movq "DCTSIZE_S"*5*2(%%"REG_D"), %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1475 "paddw %%mm1, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1476
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1477 "movq "DCTSIZE_S"*3*2(%%"REG_D"), %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1478 "paddw %%mm3, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1479
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1480 "movq 3*8+%3, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1481 "add $24, %%"REG_S" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1482
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1483 "movq %%mm7, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1484 "paddw %%mm0, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1485
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1486 "movq %%mm4, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1487 "psubw %%mm2, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1488
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1489 "movq "DCTSIZE_S"*4*2(%%"REG_D"), %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1490 "paddw %%mm2, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1491
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1492 "movq %%mm5, "DCTSIZE_S"*5*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1493 "paddw %%mm0, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1494
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1495 "movq %%mm6, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1496
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1497 "movq %%mm7, "DCTSIZE_S"*4*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1498 "add $24, %%"REG_D" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1499 "sub $2, %%"REG_c" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1500 "jnz 1b \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1501 "jmp 5f \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1502
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1503 "3: \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1504 //--- non DC2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1505 //psraw mm1, 2 w/o it -> offset. thr1, thr1, thr1 (actually thr1, thr1, thr1-1)
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1506 //psraw mm5, 2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1507 //psraw mm0, 2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1508 //psraw mm6, 2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1509 "movq %%mm5, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1510 "psubw %%mm1, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1511
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1512 "psllw $1, %%mm5 \n\t" //'z10
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1513 "paddw %%mm1, %%mm3 \n\t" //'z13
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1514
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1515 "movq %%mm0, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1516 "psubw %%mm6, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1517
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1518 "movq %%mm5, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1519 "psllw $1, %%mm0 \n\t" //'z12
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1520
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1521 "pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm1 \n\t" //-
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1522 "paddw %%mm0, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1523
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1524 "pmulhw "MANGLE(MM_FIX_1_847759065)", %%mm5 \n\t" //'z5
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1525 "paddw %%mm6, %%mm2 \n\t" //'z11
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1526
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1527 "pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1528 "movq %%mm2, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1529
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1530 //---
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1531 "movq 0*8+%3, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1532 "psubw %%mm3, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1533
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1534 "psllw $1, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1535 "paddw %%mm3, %%mm7 \n\t" //'t7
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1536
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1537 "pmulhw "MANGLE(MM_FIX_1_414213562)", %%mm2 \n\t" //'t11
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1538 "movq %%mm4, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1539 //paddw mm7, MM_2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1540 "psraw $2, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1541
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1542 "paddw "DCTSIZE_S"*0*2(%%"REG_D"), %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1543 "psubw %%mm7, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1544
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1545 "movq 1*8+%3, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1546 "paddw %%mm7, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1547
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1548 "movq %%mm6, "DCTSIZE_S"*7*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1549 "paddw %%mm5, %%mm1 \n\t" //'t12
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1550
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1551 "movq %%mm4, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1552 "psubw %%mm7, %%mm1 \n\t" //'t6
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1553
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1554 "movq 2*8+%3, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1555 "psubw %%mm5, %%mm0 \n\t" //'t10
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1556
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1557 "movq 3*8+%3, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1558 "movq %%mm3, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1559
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1560 "paddw "DCTSIZE_S"*1*2(%%"REG_D"), %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1561 "psubw %%mm1, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1562
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1563 "psubw %%mm1, %%mm2 \n\t" //'t5
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1564 "paddw %%mm1, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1565
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1566 "movq %%mm5, "DCTSIZE_S"*6*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1567 "movq %%mm7, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1568
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1569 "paddw "DCTSIZE_S"*2*2(%%"REG_D"), %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1570 "psubw %%mm2, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1571
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1572 "paddw "DCTSIZE_S"*5*2(%%"REG_D"), %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1573 "paddw %%mm2, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1574
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1575 "movq %%mm3, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1576 "paddw %%mm2, %%mm0 \n\t" //'t4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1577
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1578 // 't4 't6 't5 - - - - 't7
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1579 "movq %%mm7, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1580 "movq %%mm6, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1581
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1582 "paddw "DCTSIZE_S"*4*2(%%"REG_D"), %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1583 "psubw %%mm0, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1584
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1585 "paddw "DCTSIZE_S"*3*2(%%"REG_D"), %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1586 "paddw %%mm0, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1587
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1588 "movq %%mm4, "DCTSIZE_S"*5*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1589 "add $24, %%"REG_S" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1590
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1591 "movq %%mm6, "DCTSIZE_S"*4*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1592
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1593 "movq %%mm1, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1594 "add $24, %%"REG_D" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1595 "sub $2, %%"REG_c" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1596 "jnz 1b \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1597 "5: \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1598
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1599 : "+S"(data), "+D"(output), "+c"(cnt), "=o"(temps)
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1600 : "d"(thr_adr)
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1601 : "%"REG_a
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1602 );
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1603 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1604
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1605 #endif // HAVE_MMX
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1606
28290
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_...
reimar
parents: 27754
diff changeset
1607 #if !HAVE_MMX
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1608
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1609 static void row_idct_c(DCTELEM* workspace,
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1610 int16_t* output_adr, int output_stride, int cnt)
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1611 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1612 int_simd16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1613 int_simd16_t tmp10, tmp11, tmp12, tmp13;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1614 int_simd16_t z5, z10, z11, z12, z13;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1615 int16_t* outptr;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1616 DCTELEM* wsptr;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1617
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1618 cnt*=4;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1619 wsptr = workspace;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1620 outptr = output_adr;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1621 for (; cnt > 0; cnt--) {
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1622 // Even part
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1623 //Simd version reads 4x4 block and transposes it
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1624 tmp10 = ( wsptr[2] + wsptr[3]);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1625 tmp11 = ( wsptr[2] - wsptr[3]);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1626
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1627 tmp13 = ( wsptr[0] + wsptr[1]);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1628 tmp12 = (MULTIPLY16H( wsptr[0] - wsptr[1], FIX_1_414213562_A)<<2) - tmp13;//this shift order to avoid overflow
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1629
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1630 tmp0 = tmp10 + tmp13; //->temps
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1631 tmp3 = tmp10 - tmp13; //->temps
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1632 tmp1 = tmp11 + tmp12;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1633 tmp2 = tmp11 - tmp12;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1634
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1635 // Odd part
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1636 //Also transpose, with previous:
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1637 // ---- ---- ||||
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1638 // ---- ---- idct ||||
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1639 // ---- ---- ---> ||||
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1640 // ---- ---- ||||
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1641 z13 = wsptr[4] + wsptr[5];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1642 z10 = wsptr[4] - wsptr[5];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1643 z11 = wsptr[6] + wsptr[7];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1644 z12 = wsptr[6] - wsptr[7];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1645
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1646 tmp7 = z11 + z13;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1647 tmp11 = MULTIPLY16H(z11 - z13, FIX_1_414213562);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1648
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1649 z5 = MULTIPLY16H(z10 + z12, FIX_1_847759065);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1650 tmp10 = MULTIPLY16H(z12, FIX_1_082392200) - z5;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1651 tmp12 = MULTIPLY16H(z10, FIX_2_613125930) + z5; // - FIX_
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1652
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1653 tmp6 = (tmp12<<3) - tmp7;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1654 tmp5 = (tmp11<<3) - tmp6;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1655 tmp4 = (tmp10<<3) + tmp5;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1656
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1657 // Final output stage: descale and write column
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1658 outptr[0*output_stride]+= DESCALE(tmp0 + tmp7, 3);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1659 outptr[1*output_stride]+= DESCALE(tmp1 + tmp6, 3);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1660 outptr[2*output_stride]+= DESCALE(tmp2 + tmp5, 3);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1661 outptr[3*output_stride]+= DESCALE(tmp3 - tmp4, 3);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1662 outptr[4*output_stride]+= DESCALE(tmp3 + tmp4, 3);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1663 outptr[5*output_stride]+= DESCALE(tmp2 - tmp5, 3);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1664 outptr[6*output_stride]+= DESCALE(tmp1 - tmp6, 3); //no += ?
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1665 outptr[7*output_stride]+= DESCALE(tmp0 - tmp7, 3); //no += ?
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1666 outptr++;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1667
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1668 wsptr += DCTSIZE; // advance pointer to next row
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1669 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1670 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1671
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1672 #else /* HAVE_MMX */
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1673
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1674 static void row_idct_mmx (DCTELEM* workspace,
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1675 int16_t* output_adr, int output_stride, int cnt)
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1676 {
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1677 uint64_t __attribute__((aligned(8))) temps[4];
27754
08d18fe9da52 Change all occurrences of asm and __asm to __asm__, same as was done for FFmpeg.
diego
parents: 26727
diff changeset
1678 __asm__ volatile(
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1679 "lea (%%"REG_a",%%"REG_a",2), %%"REG_d" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1680 "1: \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1681 "movq "DCTSIZE_S"*0*2(%%"REG_S"), %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1682 //
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1683
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1684 "movq "DCTSIZE_S"*1*2(%%"REG_S"), %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1685 "movq %%mm0, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1686
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1687 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1688 "punpcklwd %%mm1, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1689
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1690 "movq "DCTSIZE_S"*3*2(%%"REG_S"), %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1691 "punpckhwd %%mm1, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1692
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1693 //transpose 4x4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1694 "movq %%mm2, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1695 "punpcklwd %%mm3, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1696
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1697 "movq %%mm0, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1698 "punpckldq %%mm2, %%mm0 \n\t" //0
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1699
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1700 "punpckhdq %%mm2, %%mm6 \n\t" //1
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1701 "movq %%mm0, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1703 "punpckhwd %%mm3, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1704 "psubw %%mm6, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1705
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1706 "pmulhw "MANGLE(MM_FIX_1_414213562_A)", %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1707 "movq %%mm4, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1708
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1709 "punpckldq %%mm7, %%mm4 \n\t" //2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1710 "paddw %%mm6, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1711
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1712 "punpckhdq %%mm7, %%mm2 \n\t" //3
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1713 "movq %%mm4, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1714
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1715 "psllw $2, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1716 "paddw %%mm2, %%mm4 \n\t" //t10
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1717
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1718 "movq "DCTSIZE_S"*0*2+"DCTSIZE_S"(%%"REG_S"), %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1719 "psubw %%mm2, %%mm1 \n\t" //t11
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1720
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1721 "movq "DCTSIZE_S"*1*2+"DCTSIZE_S"(%%"REG_S"), %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1722 "psubw %%mm5, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1723
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1724 "movq %%mm4, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1725 "paddw %%mm5, %%mm4 \n\t" //t0
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1726
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1727 "psubw %%mm5, %%mm6 \n\t" //t3
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1728 "movq %%mm1, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1729
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1730 "movq "DCTSIZE_S"*2*2+"DCTSIZE_S"(%%"REG_S"), %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1731 "paddw %%mm0, %%mm1 \n\t" //t1
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1732
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1733 "movq %%mm4, 0*8+%3 \n\t" //t0
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1734 "movq %%mm3, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1735
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1736 "movq %%mm6, 1*8+%3 \n\t" //t3
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1737 "punpcklwd %%mm2, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1738
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1739 //transpose 4x4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1740 "movq "DCTSIZE_S"*3*2+"DCTSIZE_S"(%%"REG_S"), %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1741 "punpckhwd %%mm2, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1742
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1743 "movq %%mm5, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1744 "punpcklwd %%mm6, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1745
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1746 "psubw %%mm0, %%mm7 \n\t" //t2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1747 "punpckhwd %%mm6, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1748
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1749 "movq %%mm3, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1750 "punpckldq %%mm5, %%mm3 \n\t" //4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1751
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1752 "punpckhdq %%mm5, %%mm0 \n\t" //5
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1753 "movq %%mm4, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1754
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1755 //
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1756 "movq %%mm3, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1757 "punpckldq %%mm2, %%mm4 \n\t" //6
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1758
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1759 "psubw %%mm0, %%mm3 \n\t" //z10
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1760 "punpckhdq %%mm2, %%mm5 \n\t" //7
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1761
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1762 "paddw %%mm0, %%mm6 \n\t" //z13
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1763 "movq %%mm4, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1764
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1765 "movq %%mm3, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1766 "psubw %%mm5, %%mm4 \n\t" //z12
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1767
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1768 "pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm0 \n\t" //-
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1769 "paddw %%mm4, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1770
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1771 "pmulhw "MANGLE(MM_FIX_1_847759065)", %%mm3 \n\t" //z5
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1772 "paddw %%mm5, %%mm2 \n\t" //z11 >
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1773
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1774 "pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1775 "movq %%mm2, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1776
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1777 "psubw %%mm6, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1778 "paddw %%mm6, %%mm5 \n\t" //t7
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1779
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1780 "pmulhw "MANGLE(MM_FIX_1_414213562)", %%mm2 \n\t" //t11
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1781 "paddw %%mm3, %%mm0 \n\t" //t12
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1782
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1783 "psllw $3, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1784 "psubw %%mm3, %%mm4 \n\t" //t10
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1785
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1786 "movq 0*8+%3, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1787 "movq %%mm1, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1788
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1789 "psllw $3, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1790 "psubw %%mm5, %%mm0 \n\t" //t6
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1791
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1792 "psllw $3, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1793 "paddw %%mm0, %%mm1 \n\t" //d1
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1794
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1795 "psubw %%mm0, %%mm2 \n\t" //t5
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1796 "psubw %%mm0, %%mm3 \n\t" //d6
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1797
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1798 "paddw %%mm2, %%mm4 \n\t" //t4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1799 "movq %%mm7, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1800
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1801 "paddw %%mm2, %%mm7 \n\t" //d2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1802 "psubw %%mm2, %%mm0 \n\t" //d5
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1803
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1804 "movq "MANGLE(MM_DESCALE_RND)", %%mm2 \n\t" //4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1805 "psubw %%mm5, %%mm6 \n\t" //d7
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1806
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1807 "paddw 0*8+%3, %%mm5 \n\t" //d0
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1808 "paddw %%mm2, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1809
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1810 "paddw %%mm2, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1811 "psraw $3, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1812
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1813 "paddw %%mm2, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1814 "psraw $3, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1815
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1816 "paddw (%%"REG_D"), %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1817 "psraw $3, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1818
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1819 "paddw (%%"REG_D",%%"REG_a",), %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1820 "paddw %%mm2, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1821
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1822 "paddw (%%"REG_D",%%"REG_a",2), %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1823 "paddw %%mm2, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1824
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1825 "movq %%mm5, (%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1826 "paddw %%mm2, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1827
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1828 "movq %%mm1, (%%"REG_D",%%"REG_a",) \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1829 "psraw $3, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1830
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1831 "movq %%mm7, (%%"REG_D",%%"REG_a",2) \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1832 "add %%"REG_d", %%"REG_D" \n\t" //3*ls
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1833
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1834 "movq 1*8+%3, %%mm5 \n\t" //t3
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1835 "psraw $3, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1836
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1837 "paddw (%%"REG_D",%%"REG_a",2), %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1838 "psubw %%mm4, %%mm5 \n\t" //d3
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1839
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1840 "paddw (%%"REG_D",%%"REG_d",), %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1841 "psraw $3, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1842
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1843 "paddw 1*8+%3, %%mm4 \n\t" //d4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1844 "paddw %%mm2, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1845
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1846 "paddw (%%"REG_D",%%"REG_a",4), %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1847 "paddw %%mm2, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1848
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1849 "movq %%mm0, (%%"REG_D",%%"REG_a",2) \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1850 "psraw $3, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1851
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1852 "paddw (%%"REG_D"), %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1853 "psraw $3, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1854
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1855 "paddw (%%"REG_D",%%"REG_a",), %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1856 "add $"DCTSIZE_S"*2*4, %%"REG_S" \n\t" //4 rows
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1857
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1858 "movq %%mm3, (%%"REG_D",%%"REG_d",) \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1859 "movq %%mm6, (%%"REG_D",%%"REG_a",4) \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1860 "movq %%mm5, (%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1861 "movq %%mm4, (%%"REG_D",%%"REG_a",) \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1862
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1863 "sub %%"REG_d", %%"REG_D" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1864 "add $8, %%"REG_D" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1865 "dec %%"REG_c" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1866 "jnz 1b \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1867
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1868 : "+S"(workspace), "+D"(output_adr), "+c"(cnt), "=o"(temps)
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1869 : "a"(output_stride*sizeof(short))
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1870 : "%"REG_d
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1871 );
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1872 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1873
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1874 #endif // HAVE_MMX
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1875
28290
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_...
reimar
parents: 27754
diff changeset
1876 #if !HAVE_MMX
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1877
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1878 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
1879 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1880 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
1881 int_simd16_t tmp10, tmp11, tmp12, tmp13;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1882 int_simd16_t z1, z2, z3, z4, z5, z11, z13;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1883 DCTELEM *dataptr;
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1884
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1885 cnt*=4;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1886 // Pass 1: process rows.
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1887
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1888 dataptr = data;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1889 for (; cnt > 0; cnt--) {
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1890 tmp0 = pixels[line_size*0] + pixels[line_size*7];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1891 tmp7 = pixels[line_size*0] - pixels[line_size*7];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1892 tmp1 = pixels[line_size*1] + pixels[line_size*6];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1893 tmp6 = pixels[line_size*1] - pixels[line_size*6];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1894 tmp2 = pixels[line_size*2] + pixels[line_size*5];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1895 tmp5 = pixels[line_size*2] - pixels[line_size*5];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1896 tmp3 = pixels[line_size*3] + pixels[line_size*4];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1897 tmp4 = pixels[line_size*3] - pixels[line_size*4];
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1898
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1899 // Even part
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1900
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1901 tmp10 = tmp0 + tmp3;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1902 tmp13 = tmp0 - tmp3;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1903 tmp11 = tmp1 + tmp2;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1904 tmp12 = tmp1 - tmp2;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1905 //Even columns are written first, this leads to different order of columns
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1906 //in column_fidct(), but they are processed independently, so all ok.
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1907 //Later in the row_idct() columns readed at the same order.
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1908 dataptr[2] = tmp10 + tmp11;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1909 dataptr[3] = tmp10 - tmp11;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1910
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1911 z1 = MULTIPLY16H((tmp12 + tmp13)<<2, FIX_0_707106781);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1912 dataptr[0] = tmp13 + z1;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1913 dataptr[1] = tmp13 - z1;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1914
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1915 // Odd part
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1916
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1917 tmp10 = (tmp4 + tmp5) <<2;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1918 tmp11 = (tmp5 + tmp6) <<2;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1919 tmp12 = (tmp6 + tmp7) <<2;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1920
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1921 z5 = MULTIPLY16H(tmp10 - tmp12, FIX_0_382683433);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1922 z2 = MULTIPLY16H(tmp10, FIX_0_541196100) + z5;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1923 z4 = MULTIPLY16H(tmp12, FIX_1_306562965) + z5;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1924 z3 = MULTIPLY16H(tmp11, FIX_0_707106781);
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1925
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1926 z11 = tmp7 + z3;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1927 z13 = tmp7 - z3;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1928
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1929 dataptr[4] = z13 + z2;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1930 dataptr[5] = z13 - z2;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1931 dataptr[6] = z11 + z4;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1932 dataptr[7] = z11 - z4;
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1933
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1934 pixels++; // advance pointer to next column
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1935 dataptr += DCTSIZE;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1936 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1937 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1938
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1939 #else /* HAVE_MMX */
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1940
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1941 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
1942 {
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1943 uint64_t __attribute__((aligned(8))) temps[4];
27754
08d18fe9da52 Change all occurrences of asm and __asm to __asm__, same as was done for FFmpeg.
diego
parents: 26727
diff changeset
1944 __asm__ volatile(
32702
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1945 "lea (%%"REG_a",%%"REG_a",2), %%"REG_d" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1946 "6: \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1947 "movd (%%"REG_S"), %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1948 "pxor %%mm7, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1949
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1950 "movd (%%"REG_S",%%"REG_a",), %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1951 "punpcklbw %%mm7, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1952
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1953 "movd (%%"REG_S",%%"REG_a",2), %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1954 "punpcklbw %%mm7, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1955
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1956 "punpcklbw %%mm7, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1957 "add %%"REG_d", %%"REG_S" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1958
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1959 "movq %%mm0, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1960 //
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1961
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1962 "movd (%%"REG_S",%%"REG_a",4), %%mm3 \n\t" //7 ;prefetch!
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1963 "movq %%mm1, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1964
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1965 "movd (%%"REG_S",%%"REG_d",), %%mm4 \n\t" //6
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1966 "punpcklbw %%mm7, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1967
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1968 "psubw %%mm3, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1969 "punpcklbw %%mm7, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1970
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1971 "paddw %%mm3, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1972 "psubw %%mm4, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1973
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1974 "movd (%%"REG_S",%%"REG_a",2), %%mm3 \n\t" //5
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1975 "paddw %%mm4, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1976
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1977 "movq %%mm5, 0*8+%3 \n\t" //t7
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1978 "punpcklbw %%mm7, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1979
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1980 "movq %%mm6, 1*8+%3 \n\t" //t6
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1981 "movq %%mm2, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1982
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1983 "movd (%%"REG_S"), %%mm5 \n\t" //3
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1984 "paddw %%mm3, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1985
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1986 "movd (%%"REG_S",%%"REG_a",), %%mm6 \n\t" //4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1987 "punpcklbw %%mm7, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1988
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1989 "psubw %%mm3, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1990 "punpcklbw %%mm7, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1991
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1992 "movq %%mm5, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1993 "paddw %%mm6, %%mm5 \n\t" //t3
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1994
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1995 "psubw %%mm6, %%mm3 \n\t" //t4 ; t0 t1 t2 t4 t5 t3 - -
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1996 "movq %%mm0, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1997
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1998 "movq %%mm1, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
1999 "psubw %%mm5, %%mm0 \n\t" //t13
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2000
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2001 "psubw %%mm2, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2002 "paddw %%mm2, %%mm7 \n\t" //t11
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2003
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2004 "paddw %%mm0, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2005 "movq %%mm7, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2006
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2007 "psllw $2, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2008 "paddw %%mm5, %%mm6 \n\t" //t10
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2009
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2010 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2011 "paddw %%mm6, %%mm7 \n\t" //d2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2012
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2013 "psubw %%mm2, %%mm6 \n\t" //d3
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2014 "movq %%mm0, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2015
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2016 //transpose 4x4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2017 "movq %%mm7, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2018 "punpcklwd %%mm6, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2019
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2020 "paddw %%mm1, %%mm0 \n\t" //d0
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2021 "punpckhwd %%mm6, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2022
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2023 "psubw %%mm1, %%mm5 \n\t" //d1
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2024 "movq %%mm0, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2025
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2026 "movq 1*8+%3, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2027 "punpcklwd %%mm5, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2028
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2029 "punpckhwd %%mm5, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2030 "movq %%mm0, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2031
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2032 "punpckldq %%mm7, %%mm0 \n\t" //0
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2033 "paddw %%mm4, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2034
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2035 "punpckhdq %%mm7, %%mm5 \n\t" //1
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2036 "movq %%mm6, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2037
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2038 "movq %%mm0, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2039 "punpckldq %%mm2, %%mm6 \n\t" //2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2040
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2041 "movq %%mm5, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2042 "punpckhdq %%mm2, %%mm7 \n\t" //3
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2043
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2044 "movq %%mm6, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2045 "paddw %%mm1, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2046
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2047 "movq %%mm7, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2048 "psllw $2, %%mm3 \n\t" //t10
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2049
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2050 "movq 0*8+%3, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2051 "psllw $2, %%mm4 \n\t" //t11
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2052
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2053 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm4 \n\t" //z3
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2054 "paddw %%mm2, %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2055
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2056 "psllw $2, %%mm1 \n\t" //t12
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2057 "movq %%mm3, %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2058
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2059 "pmulhw "MANGLE(MM_FIX_0_541196100)", %%mm0 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2060 "psubw %%mm1, %%mm3 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2061
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2062 "pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t" //z5
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2063 "movq %%mm2, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2064
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2065 "pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm1 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2066 "psubw %%mm4, %%mm2 \n\t" //z13
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2067
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2068 "paddw %%mm4, %%mm5 \n\t" //z11
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2069 "movq %%mm2, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2070
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2071 "paddw %%mm3, %%mm0 \n\t" //z2
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2072 "movq %%mm5, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2073
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2074 "paddw %%mm0, %%mm2 \n\t" //d4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2075 "psubw %%mm0, %%mm6 \n\t" //d5
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2076
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2077 "movq %%mm2, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2078 "paddw %%mm3, %%mm1 \n\t" //z4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2079
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2080 //transpose 4x4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2081 "punpcklwd %%mm6, %%mm2 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2082 "paddw %%mm1, %%mm5 \n\t" //d6
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2083
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2084 "punpckhwd %%mm6, %%mm4 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2085 "psubw %%mm1, %%mm7 \n\t" //d7
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2086
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2087 "movq %%mm5, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2088 "punpcklwd %%mm7, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2089
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2090 "punpckhwd %%mm7, %%mm6 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2091 "movq %%mm2, %%mm7 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2092
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2093 "punpckldq %%mm5, %%mm2 \n\t" //4
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2094 "sub %%"REG_d", %%"REG_S" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2095
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2096 "punpckhdq %%mm5, %%mm7 \n\t" //5
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2097 "movq %%mm4, %%mm5 \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2098
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2099 "movq %%mm2, "DCTSIZE_S"*0*2+"DCTSIZE_S"(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2100 "punpckldq %%mm6, %%mm4 \n\t" //6
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2101
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2102 "movq %%mm7, "DCTSIZE_S"*1*2+"DCTSIZE_S"(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2103 "punpckhdq %%mm6, %%mm5 \n\t" //7
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2104
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2105 "movq %%mm4, "DCTSIZE_S"*2*2+"DCTSIZE_S"(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2106 "add $4, %%"REG_S" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2107
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2108 "movq %%mm5, "DCTSIZE_S"*3*2+"DCTSIZE_S"(%%"REG_D") \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2109 "add $"DCTSIZE_S"*2*4, %%"REG_D" \n\t" //4 rows
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2110 "dec %%"REG_c" \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2111 "jnz 6b \n\t"
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2112
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2113 : "+S"(pixels), "+D"(data), "+c"(cnt), "=o"(temps)
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2114 : "a"(line_size)
7af3e6f901fd Convert some tabs to whitespace to allow using MPlayer filter sourcecode in FFmpeg.
cehoyos
parents: 32537
diff changeset
2115 : "%"REG_d);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2116 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2117
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2118 #endif // HAVE_MMX