annotate libmpcodecs/vf_fspp.c @ 30738:25622198a15a

Make -slave also work with the normal console on Windows.
author reimar
date Sat, 27 Feb 2010 23:12:34 +0000
parents a972c1a4a012
children f957f330aa6d
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"
17012
6ff3379a0862 Unify include path handling, -I.. is in CFLAGS.
diego
parents: 16018
diff changeset
48 #include "libvo/fastmemcpy.h"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
49
28588
7f03a6d3c941 Move FFmpeg #includes below all others so that they do not override
diego
parents: 28327
diff changeset
50 #include "libavutil/internal.h"
7f03a6d3c941 Move FFmpeg #includes below all others so that they do not override
diego
parents: 28327
diff changeset
51 #include "libavutil/intreadwrite.h"
7f03a6d3c941 Move FFmpeg #includes below all others so that they do not override
diego
parents: 28327
diff changeset
52 #include "libavutil/mem.h"
7f03a6d3c941 Move FFmpeg #includes below all others so that they do not override
diego
parents: 28327
diff changeset
53 #include "libavcodec/avcodec.h"
7f03a6d3c941 Move FFmpeg #includes below all others so that they do not override
diego
parents: 28327
diff changeset
54 #include "libavcodec/dsputil.h"
7f03a6d3c941 Move FFmpeg #includes below all others so that they do not override
diego
parents: 28327
diff changeset
55
28327
c39a1fd7d45c Fix compilation after DECLARE_ASM_CONST/DECLARE_ALIGNED moving within FFmpeg.
diego
parents: 28290
diff changeset
56 #undef free
c39a1fd7d45c Fix compilation after DECLARE_ASM_CONST/DECLARE_ALIGNED moving within FFmpeg.
diego
parents: 28290
diff changeset
57 #undef malloc
c39a1fd7d45c Fix compilation after DECLARE_ASM_CONST/DECLARE_ALIGNED moving within FFmpeg.
diego
parents: 28290
diff changeset
58
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
59 //===========================================================================//
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
60 #define BLOCKSZ 12
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 static const short custom_threshold[64]=
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
63 // values (296) can't be too high
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
64 // -it causes too big quant dependence
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
65 // or maybe overflow(check), which results in some flashing
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
66 { 71, 296, 295, 237, 71, 40, 38, 19,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
67 245, 193, 185, 121, 102, 73, 53, 27,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
68 158, 129, 141, 107, 97, 73, 50, 26,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
69 102, 116, 109, 98, 82, 66, 45, 23,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
70 71, 94, 95, 81, 70, 56, 38, 20,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
71 56, 77, 74, 66, 56, 44, 30, 15,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
72 38, 53, 50, 45, 38, 30, 21, 11,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
73 20, 27, 26, 23, 20, 15, 11, 5
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
74 };
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
75
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
76 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
77 { 0, 48, 12, 60, 3, 51, 15, 63, },
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
78 { 32, 16, 44, 28, 35, 19, 47, 31, },
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
79 { 8, 56, 4, 52, 11, 59, 7, 55, },
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
80 { 40, 24, 36, 20, 43, 27, 39, 23, },
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
81 { 2, 50, 14, 62, 1, 49, 13, 61, },
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
82 { 34, 18, 46, 30, 33, 17, 45, 29, },
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
83 { 10, 58, 6, 54, 9, 57, 5, 53, },
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
84 { 42, 26, 38, 22, 41, 25, 37, 21, },
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
85 };
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
86
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
87 struct vf_priv_s { //align 16 !
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
88 uint64_t threshold_mtx_noq[8*2];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
89 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
90
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
91 int log2_count;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
92 int temp_stride;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
93 int qp;
17225
ec9888363742 reverse the H264 hack
henry
parents: 17223
diff changeset
94 int mpeg2;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
95 int prev_q;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
96 uint8_t *src;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
97 int16_t *temp;
17225
ec9888363742 reverse the H264 hack
henry
parents: 17223
diff changeset
98 int bframes;
17133
a2b24e0d7772 prevent flicker on b-frames, trivial port from vf_spp
henry
parents: 17012
diff changeset
99 char *non_b_qp;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
100 };
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
101
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
102
28290
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_...
reimar
parents: 27754
diff changeset
103 #if !HAVE_MMX
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
104
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
105 //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
106 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
107 {int y, x;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
108 #define STORE(pos) \
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
109 temp= (src[x + pos] + (d[pos]>>log2_scale))>>(6-log2_scale); \
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
110 src[x + pos]=src[x + pos - 8*src_stride]=0; \
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
111 if(temp & 0x100) temp= ~(temp>>31); \
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
112 dst[x + pos]= temp;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
113
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
114 for(y=0; y<height; y++){
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
115 const uint8_t *d= dither[y];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
116 for(x=0; x<width; x+=8){
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
117 int temp;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
118 STORE(0);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
119 STORE(1);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
120 STORE(2);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
121 STORE(3);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
122 STORE(4);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
123 STORE(5);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
124 STORE(6);
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
125 STORE(7);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
126 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
127 src+=src_stride;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
128 dst+=dst_stride;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
129 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
130 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
131
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
132 //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
133 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
134 {int y, x;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
135 #define STORE2(pos) \
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
136 temp= (src[x + pos] + src[x + pos + 16*src_stride] + (d[pos]>>log2_scale))>>(6-log2_scale); \
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
137 src[x + pos + 16*src_stride]=0; \
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
138 if(temp & 0x100) temp= ~(temp>>31); \
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
139 dst[x + pos]= temp;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
140
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
141 for(y=0; y<height; y++){
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
142 const uint8_t *d= dither[y];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
143 for(x=0; x<width; x+=8){
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
144 int temp;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
145 STORE2(0);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
146 STORE2(1);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
147 STORE2(2);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
148 STORE2(3);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
149 STORE2(4);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
150 STORE2(5);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
151 STORE2(6);
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
152 STORE2(7);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
153 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
154 src+=src_stride;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
155 dst+=dst_stride;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
156 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
157 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
158
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
159 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
160 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
161 int a;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
162 for(a=0;a<64;a++)
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
163 ((short*)p->threshold_mtx)[a]=q * ((short*)p->threshold_mtx_noq)[a];//ints faster in C
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
164 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
165
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
166 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
167 static void row_idct_c(DCTELEM* workspace,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
168 int16_t* output_adr, int output_stride, int cnt);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
169 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
170
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
171 //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
172 #define store_slice_s store_slice_c
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
173 #define store_slice2_s store_slice2_c
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
174 #define mul_thrmat_s mul_thrmat_c
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
175 #define column_fidct_s column_fidct_c
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
176 #define row_idct_s row_idct_c
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
177 #define row_fdct_s row_fdct_c
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
178
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
179 #else /* HAVE_MMX */
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 //This func reads from 1 slice, 1 and clears 0 & 1
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
182 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
183 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
184 const uint8_t *od=&dither[0][0];
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
185 const uint8_t *end=&dither[height][0];
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
186 width = (width+7)&~7;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
187 dst_stride-=width;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
188 //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
189 __asm__ volatile(
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
190 "mov %5, %%"REG_d" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
191 "mov %6, %%"REG_S" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
192 "mov %7, %%"REG_D" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
193 "mov %1, %%"REG_a" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
194 "movd %%"REG_d", %%mm5 \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
195 "xor $-1, %%"REG_d" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
196 "mov %%"REG_a", %%"REG_c" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
197 "add $7, %%"REG_d" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
198 "neg %%"REG_a" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
199 "sub %0, %%"REG_c" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
200 "add %%"REG_c", %%"REG_c" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
201 "movd %%"REG_d", %%mm2 \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
202 "mov %%"REG_c", %1 \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
203 "mov %2, %%"REG_d" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
204 "shl $4, %%"REG_a" \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
205
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
206 "2: \n\t"
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
207 "movq (%%"REG_d"), %%mm3 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
208 "movq %%mm3, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
209 "pxor %%mm7, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
210 "punpcklbw %%mm7, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
211 "punpckhbw %%mm7, %%mm4 \n\t"
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
212 "mov %0, %%"REG_c" \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
213 "psraw %%mm5, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
214 "psraw %%mm5, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
215 "1: \n\t"
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
216 "movq %%mm7, (%%"REG_S",%%"REG_a",) \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
217 "movq (%%"REG_S"), %%mm0 \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
218 "movq 8(%%"REG_S"), %%mm1 \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
219
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
220 "movq %%mm7, 8(%%"REG_S",%%"REG_a",) \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
221 "paddw %%mm3, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
222 "paddw %%mm4, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
223
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
224 "movq %%mm7, (%%"REG_S") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
225 "psraw %%mm2, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
226 "psraw %%mm2, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
227
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
228 "movq %%mm7, 8(%%"REG_S") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
229 "packuswb %%mm1, %%mm0 \n\t"
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
230 "add $16, %%"REG_S" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
231
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
232 "movq %%mm0, (%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
233 "add $8, %%"REG_D" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
234 "sub $8, %%"REG_c" \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
235 "jg 1b \n\t"
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
236 "add %1, %%"REG_S" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
237 "add $8, %%"REG_d" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
238 "add %3, %%"REG_D" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
239 "cmp %4, %%"REG_d" \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
240 "jl 2b \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
241
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
242 :
29310
c35891e664af replace "g" asm constraint by "erm" since "g" allows 64bit immediates while
gpoirier
parents: 29263
diff changeset
243 : "m" (width), "m" (src_stride), "erm" (od), "m" (dst_stride), "erm" (end),
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
244 "m" (log2_scale), "m" (src), "m" (dst) //input
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
245 : "%"REG_a, "%"REG_c, "%"REG_d, "%"REG_S, "%"REG_D
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
246 );
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
247 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
248
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
249 //This func reads from 2 slices, 0 & 2 and clears 2-nd
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
250 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
251 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
252 const uint8_t *od=&dither[0][0];
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
253 const uint8_t *end=&dither[height][0];
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
254 width = (width+7)&~7;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
255 dst_stride-=width;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
256 //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
257 __asm__ volatile(
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
258 "mov %5, %%"REG_d" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
259 "mov %6, %%"REG_S" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
260 "mov %7, %%"REG_D" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
261 "mov %1, %%"REG_a" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
262 "movd %%"REG_d", %%mm5 \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
263 "xor $-1, %%"REG_d" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
264 "mov %%"REG_a", %%"REG_c" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
265 "add $7, %%"REG_d" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
266 "sub %0, %%"REG_c" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
267 "add %%"REG_c", %%"REG_c" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
268 "movd %%"REG_d", %%mm2 \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
269 "mov %%"REG_c", %1 \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
270 "mov %2, %%"REG_d" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
271 "shl $5, %%"REG_a" \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
272
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
273 "2: \n\t"
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
274 "movq (%%"REG_d"), %%mm3 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
275 "movq %%mm3, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
276 "pxor %%mm7, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
277 "punpcklbw %%mm7, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
278 "punpckhbw %%mm7, %%mm4 \n\t"
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
279 "mov %0, %%"REG_c" \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
280 "psraw %%mm5, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
281 "psraw %%mm5, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
282 "1: \n\t"
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
283 "movq (%%"REG_S"), %%mm0 \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
284 "movq 8(%%"REG_S"), %%mm1 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
285 "paddw %%mm3, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
286
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
287 "paddw (%%"REG_S",%%"REG_a",), %%mm0 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
288 "paddw %%mm4, %%mm1 \n\t"
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
289 "movq 8(%%"REG_S",%%"REG_a",), %%mm6 \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
290
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
291 "movq %%mm7, (%%"REG_S",%%"REG_a",) \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
292 "psraw %%mm2, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
293 "paddw %%mm6, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
294
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
295 "movq %%mm7, 8(%%"REG_S",%%"REG_a",) \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
296 "psraw %%mm2, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
297 "packuswb %%mm1, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
298
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
299 "movq %%mm0, (%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
300 "add $16, %%"REG_S" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
301 "add $8, %%"REG_D" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
302 "sub $8, %%"REG_c" \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
303 "jg 1b \n\t"
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
304 "add %1, %%"REG_S" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
305 "add $8, %%"REG_d" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
306 "add %3, %%"REG_D" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
307 "cmp %4, %%"REG_d" \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
308 "jl 2b \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
309
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
310 :
29310
c35891e664af replace "g" asm constraint by "erm" since "g" allows 64bit immediates while
gpoirier
parents: 29263
diff changeset
311 : "m" (width), "m" (src_stride), "erm" (od), "m" (dst_stride), "erm" (end),
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
312 "m" (log2_scale), "m" (src), "m" (dst) //input
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
313 : "%"REG_a, "%"REG_c, "%"REG_d, "%"REG_D, "%"REG_S
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
314 );
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
315 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
316
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
317 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
318 {
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
319 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
320 __asm__ volatile(
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
321 "movd %0, %%mm7 \n\t"
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
322 "add $8*8*2, %%"REG_D" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
323 "movq 0*8(%%"REG_S"), %%mm0 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
324 "punpcklwd %%mm7, %%mm7 \n\t"
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
325 "movq 1*8(%%"REG_S"), %%mm1 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
326 "punpckldq %%mm7, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
327 "pmullw %%mm7, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
328
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
329 "movq 2*8(%%"REG_S"), %%mm2 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
330 "pmullw %%mm7, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
331
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
332 "movq 3*8(%%"REG_S"), %%mm3 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
333 "pmullw %%mm7, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
334
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
335 "movq %%mm0, 0*8(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
336 "movq 4*8(%%"REG_S"), %%mm4 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
337 "pmullw %%mm7, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
338
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
339 "movq %%mm1, 1*8(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
340 "movq 5*8(%%"REG_S"), %%mm5 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
341 "pmullw %%mm7, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
342
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
343 "movq %%mm2, 2*8(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
344 "movq 6*8(%%"REG_S"), %%mm6 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
345 "pmullw %%mm7, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
346
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
347 "movq %%mm3, 3*8(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
348 "movq 7*8+0*8(%%"REG_S"), %%mm0 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
349 "pmullw %%mm7, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
350
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
351 "movq %%mm4, 4*8(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
352 "movq 7*8+1*8(%%"REG_S"), %%mm1 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
353 "pmullw %%mm7, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
354
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
355 "movq %%mm5, 5*8(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
356 "movq 7*8+2*8(%%"REG_S"), %%mm2 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
357 "pmullw %%mm7, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
358
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
359 "movq %%mm6, 6*8(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
360 "movq 7*8+3*8(%%"REG_S"), %%mm3 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
361 "pmullw %%mm7, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
362
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
363 "movq %%mm0, 7*8+0*8(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
364 "movq 7*8+4*8(%%"REG_S"), %%mm4 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
365 "pmullw %%mm7, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
366
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
367 "movq %%mm1, 7*8+1*8(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
368 "movq 7*8+5*8(%%"REG_S"), %%mm5 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
369 "pmullw %%mm7, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
370
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
371 "movq %%mm2, 7*8+2*8(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
372 "movq 7*8+6*8(%%"REG_S"), %%mm6 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
373 "pmullw %%mm7, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
374
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
375 "movq %%mm3, 7*8+3*8(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
376 "movq 14*8+0*8(%%"REG_S"), %%mm0 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
377 "pmullw %%mm7, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
378
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
379 "movq %%mm4, 7*8+4*8(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
380 "movq 14*8+1*8(%%"REG_S"), %%mm1 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
381 "pmullw %%mm7, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
382
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
383 "movq %%mm5, 7*8+5*8(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
384 "pmullw %%mm7, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
385
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
386 "movq %%mm6, 7*8+6*8(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
387 "movq %%mm0, 14*8+0*8(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
388 "movq %%mm1, 14*8+1*8(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
389
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
390 : "+g" (q), "+S" (adr), "+D" (adr)
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
391 :
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
392 );
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
393 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
394
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
395 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
396 static void row_idct_mmx(DCTELEM* workspace,
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
397 int16_t* output_adr, int output_stride, int cnt);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
398 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
399
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
400 #define store_slice_s store_slice_mmx
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
401 #define store_slice2_s store_slice2_mmx
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
402 #define mul_thrmat_s mul_thrmat_mmx
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
403 #define column_fidct_s column_fidct_mmx
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
404 #define row_idct_s row_idct_mmx
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
405 #define row_fdct_s row_fdct_mmx
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
406 #endif // HAVE_MMX
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
407
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
408 static void filter(struct vf_priv_s *p, uint8_t *dst, uint8_t *src,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
409 int dst_stride, int src_stride,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
410 int width, int height,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
411 uint8_t *qp_store, int qp_stride, int is_luma)
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
412 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
413 int x, x0, y, es, qy, t;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
414 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
415 const int step=6-p->log2_count;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
416 const int qps= 3 + is_luma;
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
417 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
418 DCTELEM *block= (DCTELEM *)block_align;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
419 DCTELEM *block3=(DCTELEM *)(block_align+4*8*BLOCKSZ);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
420
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
421 memset(block3, 0, 4*8*BLOCKSZ);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
422
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
423 //p->src=src-src_stride*8-8;//!
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
424 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
425 for(y=0; y<height; y++){
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
426 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
427 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
428 for(x=0; x<8; x++){
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
429 p->src[index - x - 1]= p->src[index + x ];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
430 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
431 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
432 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
433 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
434 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
435 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
436 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
437 //FIXME (try edge emu)
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 for(y=8; y<24; y++)
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
440 memset(p->temp+ 8 +y*stride, 0,width*sizeof(int16_t));
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
441
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
442 for(y=step; y<height+8; y+=step){ //step= 1,2
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
443 qy=y-4;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
444 if (qy>height-1) qy=height-1;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
445 if (qy<0) qy=0;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
446 qy=(qy>>qps)*qp_stride;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
447 row_fdct_s(block, p->src + y*stride +2-(y&1), stride, 2);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
448 for(x0=0; x0<width+8-8*(BLOCKSZ-1); x0+=8*(BLOCKSZ-1)){
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
449 row_fdct_s(block+8*8, p->src + y*stride+8+x0 +2-(y&1), stride, 2*(BLOCKSZ-1));
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
450 if(p->qp)
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
451 column_fidct_s((int16_t*)(&p->threshold_mtx[0]), block+0*8, block3+0*8, 8*(BLOCKSZ-1)); //yes, this is a HOTSPOT
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
452 else
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
453 for (x=0; x<8*(BLOCKSZ-1); x+=8) {
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
454 t=x+x0-2; //correct t=x+x0-2-(y&1), but its the same
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
455 if (t<0) t=0;//t always < width-2
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
456 t=qp_store[qy+(t>>qps)];
30412
41fb4acf3df6 Support more qscale types in most post-processing filters.
reimar
parents: 30363
diff changeset
457 t=norm_qscale(t, p->mpeg2);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
458 if (t!=p->prev_q) p->prev_q=t, mul_thrmat_s(p, t);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
459 column_fidct_s((int16_t*)(&p->threshold_mtx[0]), block+x*8, block3+x*8, 8); //yes, this is a HOTSPOT
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
460 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
461 row_idct_s(block3+0*8, p->temp + (y&15)*stride+x0+2-(y&1), stride, 2*(BLOCKSZ-1));
25568
707b810a2558 fix artifacts in -vf fspp. regression in r23476.
lorenm
parents: 25221
diff changeset
462 memmove(block, block+(BLOCKSZ-1)*64, 8*8*sizeof(DCTELEM)); //cycling
707b810a2558 fix artifacts in -vf fspp. regression in r23476.
lorenm
parents: 25221
diff changeset
463 memmove(block3, block3+(BLOCKSZ-1)*64, 6*8*sizeof(DCTELEM));
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
464 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
465 //
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
466 es=width+8-x0; // 8, ...
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
467 if (es>8)
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
468 row_fdct_s(block+8*8, p->src + y*stride+8+x0 +2-(y&1), stride, (es-4)>>2);
21578
9345eb2d8c8f count needs to be even
henry
parents: 20585
diff changeset
469 column_fidct_s((int16_t*)(&p->threshold_mtx[0]), block, block3, es&(~1));
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
470 row_idct_s(block3+0*8, p->temp + (y&15)*stride+x0+2-(y&1), stride, es>>2);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
471 {const int y1=y-8+step;//l5-7 l4-6
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
472 if (!(y1&7) && y1) {
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
473 if (y1&8) store_slice_s(dst + (y1-8)*dst_stride, p->temp+ 8 +8*stride,
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
474 dst_stride, stride, width, 8, 5-p->log2_count);
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
475 else store_slice2_s(dst + (y1-8)*dst_stride, p->temp+ 8 +0*stride,
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
476 dst_stride, stride, width, 8, 5-p->log2_count);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
477 } }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
478 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
479
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
480 if (y&7) { // == height & 7
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
481 if (y&8) store_slice_s(dst + ((y-8)&~7)*dst_stride, p->temp+ 8 +8*stride,
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
482 dst_stride, stride, width, y&7, 5-p->log2_count);
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
483 else store_slice2_s(dst + ((y-8)&~7)*dst_stride, p->temp+ 8 +0*stride,
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
484 dst_stride, stride, width, y&7, 5-p->log2_count);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
485 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
486 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
487
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
488 static int config(struct vf_instance *vf,
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
489 int width, int height, int d_width, int d_height,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
490 unsigned int flags, unsigned int outfmt)
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
491 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
492 int h= (height+16+15)&(~15);
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 vf->priv->temp_stride= (width+16+15)&(~15);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
495 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
496 //this can also be avoided, see above
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
497 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
498
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
499 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
500 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
501
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
502 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
503 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
504 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
505 // 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
506 vf->dmpi=vf_get_image(vf->next,mpi->imgfmt,
16018
bdf1b4ecb906 use stored dimensions instead of visible one when (vf_)get_image is called
iive
parents: 15651
diff changeset
507 mpi->type, mpi->flags, mpi->width, mpi->height);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
508 mpi->planes[0]=vf->dmpi->planes[0];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
509 mpi->stride[0]=vf->dmpi->stride[0];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
510 mpi->width=vf->dmpi->width;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
511 if(mpi->flags&MP_IMGFLAG_PLANAR){
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
512 mpi->planes[1]=vf->dmpi->planes[1];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
513 mpi->planes[2]=vf->dmpi->planes[2];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
514 mpi->stride[1]=vf->dmpi->stride[1];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
515 mpi->stride[2]=vf->dmpi->stride[2];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
516 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
517 mpi->flags|=MP_IMGFLAG_DIRECT;
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
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
520 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
521 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
522 mp_image_t *dmpi;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
523 if(!(mpi->flags&MP_IMGFLAG_DIRECT)){
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
524 // no DR, so get a new image! hope we'll get DR buffer:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
525 dmpi=vf_get_image(vf->next,mpi->imgfmt,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
526 MP_IMGTYPE_TEMP,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
527 MP_IMGFLAG_ACCEPT_STRIDE|MP_IMGFLAG_PREFER_ALIGNED_STRIDE,
16018
bdf1b4ecb906 use stored dimensions instead of visible one when (vf_)get_image is called
iive
parents: 15651
diff changeset
528 mpi->width,mpi->height);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
529 vf_clone_mpi_attributes(dmpi, mpi);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
530 }else{
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
531 dmpi=vf->dmpi;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
532 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
533
17225
ec9888363742 reverse the H264 hack
henry
parents: 17223
diff changeset
534 vf->priv->mpeg2= mpi->qscale_type;
17133
a2b24e0d7772 prevent flicker on b-frames, trivial port from vf_spp
henry
parents: 17012
diff changeset
535 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
536 int w = mpi->qstride;
915be5c7a30c Make sure that a qstride of 0 (intentional or not) does not completely break
reimar
parents: 29310
diff changeset
537 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
538 if (!w) {
915be5c7a30c Make sure that a qstride of 0 (intentional or not) does not completely break
reimar
parents: 29310
diff changeset
539 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
540 h = 1;
915be5c7a30c Make sure that a qstride of 0 (intentional or not) does not completely break
reimar
parents: 29310
diff changeset
541 }
915be5c7a30c Make sure that a qstride of 0 (intentional or not) does not completely break
reimar
parents: 29310
diff changeset
542 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
543 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
544 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
545 }
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
546 if(vf->priv->log2_count || !(mpi->flags&MP_IMGFLAG_DIRECT)){
17133
a2b24e0d7772 prevent flicker on b-frames, trivial port from vf_spp
henry
parents: 17012
diff changeset
547 char *qp_tab= vf->priv->non_b_qp;
17225
ec9888363742 reverse the H264 hack
henry
parents: 17223
diff changeset
548 if(vf->priv->bframes || !qp_tab)
17133
a2b24e0d7772 prevent flicker on b-frames, trivial port from vf_spp
henry
parents: 17012
diff changeset
549 qp_tab= mpi->qscale;
a2b24e0d7772 prevent flicker on b-frames, trivial port from vf_spp
henry
parents: 17012
diff changeset
550
17225
ec9888363742 reverse the H264 hack
henry
parents: 17223
diff changeset
551 if(qp_tab || vf->priv->qp){
17133
a2b24e0d7772 prevent flicker on b-frames, trivial port from vf_spp
henry
parents: 17012
diff changeset
552 filter(vf->priv, dmpi->planes[0], mpi->planes[0], dmpi->stride[0], mpi->stride[0],
a2b24e0d7772 prevent flicker on b-frames, trivial port from vf_spp
henry
parents: 17012
diff changeset
553 mpi->w, mpi->h, qp_tab, mpi->qstride, 1);
a2b24e0d7772 prevent flicker on b-frames, trivial port from vf_spp
henry
parents: 17012
diff changeset
554 filter(vf->priv, dmpi->planes[1], mpi->planes[1], dmpi->stride[1], mpi->stride[1],
a2b24e0d7772 prevent flicker on b-frames, trivial port from vf_spp
henry
parents: 17012
diff changeset
555 mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, qp_tab, mpi->qstride, 0);
a2b24e0d7772 prevent flicker on b-frames, trivial port from vf_spp
henry
parents: 17012
diff changeset
556 filter(vf->priv, dmpi->planes[2], mpi->planes[2], dmpi->stride[2], mpi->stride[2],
a2b24e0d7772 prevent flicker on b-frames, trivial port from vf_spp
henry
parents: 17012
diff changeset
557 mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, qp_tab, mpi->qstride, 0);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
558 }else{
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
559 memcpy_pic(dmpi->planes[0], mpi->planes[0], mpi->w, mpi->h, dmpi->stride[0], mpi->stride[0]);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
560 memcpy_pic(dmpi->planes[1], mpi->planes[1], mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, dmpi->stride[1], mpi->stride[1]);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
561 memcpy_pic(dmpi->planes[2], mpi->planes[2], mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, dmpi->stride[2], mpi->stride[2]);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
562 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
563 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
564
28290
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_...
reimar
parents: 27754
diff changeset
565 #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
566 if(gCpuCaps.hasMMX) __asm__ volatile ("emms\n\t");
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
567 #endif
28290
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_...
reimar
parents: 27754
diff changeset
568 #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
569 if(gCpuCaps.hasMMX2) __asm__ volatile ("sfence\n\t");
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
570 #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
571 return vf_next_put_image(vf,dmpi, pts);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
572 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
573
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
574 static void uninit(struct vf_instance *vf)
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
575 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
576 if(!vf->priv) return;
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->temp) av_free(vf->priv->temp);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
579 vf->priv->temp= NULL;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
580 if(vf->priv->src) av_free(vf->priv->src);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
581 vf->priv->src= NULL;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
582 //if(vf->priv->avctx) free(vf->priv->avctx);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
583 //vf->priv->avctx= NULL;
17133
a2b24e0d7772 prevent flicker on b-frames, trivial port from vf_spp
henry
parents: 17012
diff changeset
584 if(vf->priv->non_b_qp) free(vf->priv->non_b_qp);
a2b24e0d7772 prevent flicker on b-frames, trivial port from vf_spp
henry
parents: 17012
diff changeset
585 vf->priv->non_b_qp= NULL;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
586
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
587 av_free(vf->priv);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
588 vf->priv=NULL;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
589 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
590
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
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
593 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
594 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
595 switch(fmt){
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
596 case IMGFMT_YVU9:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
597 case IMGFMT_IF09:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
598 case IMGFMT_YV12:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
599 case IMGFMT_I420:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
600 case IMGFMT_IYUV:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
601 case IMGFMT_CLPL:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
602 case IMGFMT_Y800:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
603 case IMGFMT_Y8:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
604 case IMGFMT_444P:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
605 case IMGFMT_422P:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
606 case IMGFMT_411P:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
607 return vf_next_query_format(vf,fmt);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
608 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
609 return 0;
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
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
612 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
613 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
614 switch(request){
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
615 case VFCTRL_QUERY_MAX_PP_LEVEL:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
616 return 5;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
617 case VFCTRL_SET_PP_LEVEL:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
618 vf->priv->log2_count= *((unsigned int*)data);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
619 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
620 return CONTROL_TRUE;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
621 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
622 return vf_next_control(vf,request,data);
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
30638
a7b908875c14 Rename open() vf initialization function to vf_open().
diego
parents: 30412
diff changeset
625 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
626 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
627 int i=0, bias;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
628 int custom_threshold_m[64];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
629 int log2c=-1;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
630
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
631 vf->config=config;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
632 vf->put_image=put_image;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
633 vf->get_image=get_image;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
634 vf->query_format=query_format;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
635 vf->uninit=uninit;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
636 vf->control= control;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
637 vf->priv=av_mallocz(sizeof(struct vf_priv_s));//assumes align 16 !
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
638
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
639 avcodec_init();
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
640
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
641 //vf->priv->avctx= avcodec_alloc_context();
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
642 //dsputil_init(&vf->priv->dsp, vf->priv->avctx);
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
643
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
644 vf->priv->log2_count= 4;
17225
ec9888363742 reverse the H264 hack
henry
parents: 17223
diff changeset
645 vf->priv->bframes = 0;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
646
17225
ec9888363742 reverse the H264 hack
henry
parents: 17223
diff changeset
647 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
648
ec9888363742 reverse the H264 hack
henry
parents: 17223
diff changeset
649 if( log2c >=4 && log2c <=5 )
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
650 vf->priv->log2_count = log2c;
15651
6a0494e09435 sanity checks for options; treat quality > 5 as 5, not 4
henry
parents: 15634
diff changeset
651 else if( log2c >= 6 )
6a0494e09435 sanity checks for options; treat quality > 5 as 5, not 4
henry
parents: 15634
diff changeset
652 vf->priv->log2_count = 5;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
653
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
654 if(vf->priv->qp < 0)
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
655 vf->priv->qp = 0;
15651
6a0494e09435 sanity checks for options; treat quality > 5 as 5, not 4
henry
parents: 15634
diff changeset
656
6a0494e09435 sanity checks for options; treat quality > 5 as 5, not 4
henry
parents: 15634
diff changeset
657 if (i < -15) i = -15;
6a0494e09435 sanity checks for options; treat quality > 5 as 5, not 4
henry
parents: 15634
diff changeset
658 if (i > 32) i = 32;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
659
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
660 bias= (1<<4)+i; //regulable
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
661 vf->priv->prev_q=0;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
662 //
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
663 for(i=0;i<64;i++) //FIXME: tune custom_threshold[] and remove this !
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
664 custom_threshold_m[i]=(int)(custom_threshold[i]*(bias/71.)+ 0.5);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
665 for(i=0;i<8;i++){
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
666 vf->priv->threshold_mtx_noq[2*i]=(uint64_t)custom_threshold_m[i*8+2]
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
667 |(((uint64_t)custom_threshold_m[i*8+6])<<16)
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
668 |(((uint64_t)custom_threshold_m[i*8+0])<<32)
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
669 |(((uint64_t)custom_threshold_m[i*8+4])<<48);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
670 vf->priv->threshold_mtx_noq[2*i+1]=(uint64_t)custom_threshold_m[i*8+5]
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
671 |(((uint64_t)custom_threshold_m[i*8+3])<<16)
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
672 |(((uint64_t)custom_threshold_m[i*8+1])<<32)
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
673 |(((uint64_t)custom_threshold_m[i*8+7])<<48);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
674 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
675
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
676 if (vf->priv->qp) vf->priv->prev_q=vf->priv->qp, mul_thrmat_s(vf->priv, vf->priv->qp);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
677
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
678 return 1;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
679 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
680
25221
00fff9a3b735 Make all vf_info_t structs const
reimar
parents: 24976
diff changeset
681 const vf_info_t vf_info_fspp = {
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
682 "fast simple postprocess",
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
683 "fspp",
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
684 "Michael Niedermayer, Nikolaj Poroshin",
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
685 "",
30638
a7b908875c14 Rename open() vf initialization function to vf_open().
diego
parents: 30412
diff changeset
686 vf_open,
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
687 NULL
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
688 };
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
689
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
690 //====================================================================
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
691 //Specific spp's dct, idct and threshold functions
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
692 //I'd prefer to have them in the separate file.
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
693
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
694 //#define MANGLE(a) #a
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 //typedef int16_t DCTELEM; //! only int16_t
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 #define DCTSIZE 8
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
699 #define DCTSIZE_S "8"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
700
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
701 #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
702 #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
703 #define FIX64(x,s) C64(FIX(x,s))
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
704
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
705 #define MULTIPLY16H(x,k) (((x)*(k))>>16)
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
706 #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
707 #define DESCALE(x,n) (((x) + (1 << ((n)-1))) >> n)
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
708
28290
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_...
reimar
parents: 27754
diff changeset
709 #if HAVE_MMX
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
710
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
711 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
712 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
713 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
714 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
715
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
716 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
717
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
718 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
719 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
720 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
721 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
722 //for t3,t5,t7 == 0 shortcut
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
723 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
724 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
725 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
726
25901
c2210e68a2a9 Simplify: use DECLARE_ASM_CONST
reimar
parents: 25568
diff changeset
727 DECLARE_ASM_CONST(8, uint64_t, MM_DESCALE_RND)=C64(4);
c2210e68a2a9 Simplify: use DECLARE_ASM_CONST
reimar
parents: 25568
diff changeset
728 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
729
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
730 #else /* !HAVE_MMX */
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 typedef int32_t int_simd16_t;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
733 static const int16_t FIX_0_382683433=FIX(0.382683433, 14);
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
734 static const int16_t FIX_0_541196100=FIX(0.541196100, 14);
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
735 static const int16_t FIX_0_707106781=FIX(0.707106781, 14);
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
736 static const int16_t FIX_1_306562965=FIX(1.306562965, 14);
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
737 static const int16_t FIX_1_414213562_A=FIX(1.414213562, 14);
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
738 static const int16_t FIX_1_847759065=FIX(1.847759065, 13);
25902
15ab840747e2 mark constants as such
reimar
parents: 25901
diff changeset
739 static const int16_t FIX_2_613125930=FIX(-2.613125930, 13); //-
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
740 static const int16_t FIX_1_414213562=FIX(1.414213562, 13);
25902
15ab840747e2 mark constants as such
reimar
parents: 25901
diff changeset
741 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
742
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
743 #endif
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
744
28290
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_...
reimar
parents: 27754
diff changeset
745 #if !HAVE_MMX
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
746
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
747 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
748 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
749 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
750 int_simd16_t tmp10, tmp11, tmp12, tmp13;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
751 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
752 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
753
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
754 DCTELEM* dataptr;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
755 DCTELEM* wsptr;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
756 int16_t *threshold;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
757 int ctr;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
758
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
759 dataptr = data;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
760 wsptr = output;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
761
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
762 for (; cnt > 0; cnt-=2) { //start positions
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
763 threshold=(int16_t*)thr_adr;//threshold_mtx
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
764 for (ctr = DCTSIZE; ctr > 0; ctr--) {
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
765 // Process columns from input, add to output.
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
766 tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
767 tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
768
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
769 tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
770 tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
771
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
772 tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
773 tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
774
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
775 tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
776 tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
777
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
778 // Even part of FDCT
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
779
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
780 tmp10 = tmp0 + tmp3;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
781 tmp13 = tmp0 - tmp3;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
782 tmp11 = tmp1 + tmp2;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
783 tmp12 = tmp1 - tmp2;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
784
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
785 d0 = tmp10 + tmp11;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
786 d4 = tmp10 - tmp11;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
787
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
788 z1 = MULTIPLY16H((tmp12 + tmp13) <<2, FIX_0_707106781);
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
789 d2 = tmp13 + z1;
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
790 d6 = tmp13 - z1;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
791
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
792 // Even part of IDCT
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
793
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
794 THRESHOLD(tmp0, d0, threshold[0*8]);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
795 THRESHOLD(tmp1, d2, threshold[2*8]);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
796 THRESHOLD(tmp2, d4, threshold[4*8]);
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
797 THRESHOLD(tmp3, d6, threshold[6*8]);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
798 tmp0+=2;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
799 tmp10 = (tmp0 + tmp2)>>2;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
800 tmp11 = (tmp0 - tmp2)>>2;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
801
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
802 tmp13 = (tmp1 + tmp3)>>2; //+2 ! (psnr decides)
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
803 tmp12 = MULTIPLY16H((tmp1 - tmp3), FIX_1_414213562_A) - tmp13; //<<2
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
804
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
805 tmp0 = tmp10 + tmp13; //->temps
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
806 tmp3 = tmp10 - tmp13; //->temps
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
807 tmp1 = tmp11 + tmp12; //->temps
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
808 tmp2 = tmp11 - tmp12; //->temps
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
809
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
810 // Odd part of FDCT
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
811
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
812 tmp10 = tmp4 + tmp5;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
813 tmp11 = tmp5 + tmp6;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
814 tmp12 = tmp6 + tmp7;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
815
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
816 z5 = MULTIPLY16H((tmp10 - tmp12)<<2, FIX_0_382683433);
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
817 z2 = MULTIPLY16H(tmp10 <<2, FIX_0_541196100) + z5;
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
818 z4 = MULTIPLY16H(tmp12 <<2, FIX_1_306562965) + z5;
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
819 z3 = MULTIPLY16H(tmp11 <<2, FIX_0_707106781);
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
820
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
821 z11 = tmp7 + z3;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
822 z13 = tmp7 - z3;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
823
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
824 d5 = z13 + z2;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
825 d3 = z13 - z2;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
826 d1 = z11 + z4;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
827 d7 = z11 - z4;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
828
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
829 // Odd part of IDCT
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
830
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
831 THRESHOLD(tmp4, d1, threshold[1*8]);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
832 THRESHOLD(tmp5, d3, threshold[3*8]);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
833 THRESHOLD(tmp6, d5, threshold[5*8]);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
834 THRESHOLD(tmp7, d7, threshold[7*8]);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
835
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
836 //Simd version uses here a shortcut for the tmp5,tmp6,tmp7 == 0
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
837 z13 = tmp6 + tmp5;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
838 z10 = (tmp6 - tmp5)<<1;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
839 z11 = tmp4 + tmp7;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
840 z12 = (tmp4 - tmp7)<<1;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
841
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
842 tmp7 = (z11 + z13)>>2; //+2 !
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
843 tmp11 = MULTIPLY16H((z11 - z13)<<1, FIX_1_414213562);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
844 z5 = MULTIPLY16H(z10 + z12, FIX_1_847759065);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
845 tmp10 = MULTIPLY16H(z12, FIX_1_082392200) - z5;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
846 tmp12 = MULTIPLY16H(z10, FIX_2_613125930) + z5; // - !!
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
847
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
848 tmp6 = tmp12 - tmp7;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
849 tmp5 = tmp11 - tmp6;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
850 tmp4 = tmp10 + tmp5;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
851
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
852 wsptr[DCTSIZE*0]+= (tmp0 + tmp7);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
853 wsptr[DCTSIZE*1]+= (tmp1 + tmp6);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
854 wsptr[DCTSIZE*2]+= (tmp2 + tmp5);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
855 wsptr[DCTSIZE*3]+= (tmp3 - tmp4);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
856 wsptr[DCTSIZE*4]+= (tmp3 + tmp4);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
857 wsptr[DCTSIZE*5]+= (tmp2 - tmp5);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
858 wsptr[DCTSIZE*6]= (tmp1 - tmp6);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
859 wsptr[DCTSIZE*7]= (tmp0 - tmp7);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
860 //
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
861 dataptr++; //next column
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
862 wsptr++;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
863 threshold++;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
864 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
865 dataptr+=8; //skip each second start pos
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
866 wsptr +=8;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
867 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
868 }
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 #else /* HAVE_MMX */
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 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
873 {
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
874 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
875 __asm__ volatile(
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18131
diff changeset
876 ASMALIGN(4)
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
877 "1: \n\t"
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
878 "movq "DCTSIZE_S"*0*2(%%"REG_S"), %%mm1 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
879 //
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
880 "movq "DCTSIZE_S"*3*2(%%"REG_S"), %%mm7 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
881 "movq %%mm1, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
882
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
883 "paddw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm1 \n\t" //t0
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
884 "movq %%mm7, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
885
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
886 "paddw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm7 \n\t" //t3
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
887 "movq %%mm1, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
888
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
889 "movq "DCTSIZE_S"*1*2(%%"REG_S"), %%mm6 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
890 "psubw %%mm7, %%mm1 \n\t" //t13
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
891
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
892 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
893 "movq %%mm6, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
894
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
895 "paddw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm6 \n\t" //t1
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
896 "paddw %%mm7, %%mm5 \n\t" //t10
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
897
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
898 "paddw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t2
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
899 "movq %%mm6, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
900
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
901 "paddw %%mm2, %%mm6 \n\t" //t11
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
902 "psubw %%mm2, %%mm7 \n\t" //t12
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
903
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
904 "movq %%mm5, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
905 "paddw %%mm6, %%mm5 \n\t" //d0
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
906 // i0 t13 t12 i3 i1 d0 - d4
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
907 "psubw %%mm6, %%mm2 \n\t" //d4
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
908 "paddw %%mm1, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
909
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
910 "movq 4*16(%%"REG_d"), %%mm6 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
911 "psllw $2, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
912
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
913 "psubw 0*16(%%"REG_d"), %%mm5 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
914 "psubw %%mm6, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
915
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
916 "paddusw 0*16(%%"REG_d"), %%mm5 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
917 "paddusw %%mm6, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
918
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
919 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
920 //
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
921 "paddw 0*16(%%"REG_d"), %%mm5 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
922 "paddw %%mm6, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
923
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
924 "psubusw 0*16(%%"REG_d"), %%mm5 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
925 "psubusw %%mm6, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
926
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
927 //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
928 // 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
929 //However, typical numbers: nondc - 29%%, dc - 46%%, zero - 25%%. All <> 0 case is very rare.
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
930 "paddw "MANGLE(MM_2)", %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
931 "movq %%mm2, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
932
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
933 "paddw %%mm5, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
934 "psubw %%mm6, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
935
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
936 "movq %%mm1, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
937 "paddw %%mm7, %%mm1 \n\t" //d2
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
938
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
939 "psubw 2*16(%%"REG_d"), %%mm1 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
940 "psubw %%mm7, %%mm6 \n\t" //d6
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
941
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
942 "movq 6*16(%%"REG_d"), %%mm7 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
943 "psraw $2, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
944
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
945 "paddusw 2*16(%%"REG_d"), %%mm1 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
946 "psubw %%mm7, %%mm6 \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
947 // t7 d2 /t11 t4 t6 - d6 /t10
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
948
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
949 "paddw 2*16(%%"REG_d"), %%mm1 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
950 "paddusw %%mm7, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
951
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
952 "psubusw 2*16(%%"REG_d"), %%mm1 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
953 "paddw %%mm7, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
954
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
955 "psubw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm3 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
956 "psubusw %%mm7, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
957
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
958 //movq [edi+"DCTSIZE_S"*2*2], mm1
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
959 //movq [edi+"DCTSIZE_S"*6*2], mm6
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
960 "movq %%mm1, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
961 "psraw $2, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
962
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
963 "psubw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm4 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
964 "psubw %%mm6, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
965
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
966 "psubw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm0 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
967 "paddw %%mm7, %%mm6 \n\t" //'t13
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
968
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
969 "psraw $2, %%mm6 \n\t" //paddw mm6, MM_2 !! ---
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
970 "movq %%mm2, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
971
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
972 "pmulhw "MANGLE(MM_FIX_1_414213562_A)", %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
973 "paddw %%mm6, %%mm2 \n\t" //'t0
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
974
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
975 "movq %%mm2, 0*8+%3 \n\t" //!
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
976 "psubw %%mm6, %%mm7 \n\t" //'t3
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
977
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
978 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
979 "psubw %%mm6, %%mm1 \n\t" //'t12
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
980
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
981 "psubw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t5
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
982 "movq %%mm5, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
983
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
984 "movq %%mm7, 3*8+%3 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
985 "paddw %%mm2, %%mm3 \n\t" //t10
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
986
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
987 "paddw %%mm4, %%mm2 \n\t" //t11
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
988 "paddw %%mm0, %%mm4 \n\t" //t12
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
989
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
990 "movq %%mm3, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
991 "psubw %%mm4, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
992
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
993 "psllw $2, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
994 "psllw $2, %%mm7 \n\t" //opt for P6
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
995
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
996 "pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
997 "psllw $2, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
998
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
999 "pmulhw "MANGLE(MM_FIX_0_541196100)", %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1000 "psllw $2, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1001
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1002 "pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1003 "paddw %%mm1, %%mm5 \n\t" //'t1
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1004
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1005 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1006 "psubw %%mm1, %%mm6 \n\t" //'t2
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1007 // t7 't12 't11 t4 t6 - 't13 't10 ---
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1008
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1009 "paddw %%mm3, %%mm7 \n\t" //z2
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1010
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1011 "movq %%mm5, 1*8+%3 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1012 "paddw %%mm3, %%mm4 \n\t" //z4
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1013
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1014 "movq 3*16(%%"REG_d"), %%mm3 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1015 "movq %%mm0, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1016
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1017 "movq %%mm6, 2*8+%3 \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1018 "psubw %%mm2, %%mm1 \n\t" //z13
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1019
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1020 //===
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1021 "paddw %%mm2, %%mm0 \n\t" //z11
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1022 "movq %%mm1, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1023
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1024 "movq 5*16(%%"REG_d"), %%mm2 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1025 "psubw %%mm7, %%mm1 \n\t" //d3
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1026
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1027 "paddw %%mm7, %%mm5 \n\t" //d5
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1028 "psubw %%mm3, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1029
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1030 "movq 1*16(%%"REG_d"), %%mm7 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1031 "psubw %%mm2, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1032
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1033 "movq %%mm0, %%mm6 \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1034 "paddw %%mm4, %%mm0 \n\t" //d1
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1035
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1036 "paddusw %%mm3, %%mm1 \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1037 "psubw %%mm4, %%mm6 \n\t" //d7
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1038
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1039 // d1 d3 - - - d5 d7 -
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1040 "movq 7*16(%%"REG_d"), %%mm4 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1041 "psubw %%mm7, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1042
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1043 "psubw %%mm4, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1044 "paddusw %%mm2, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1045
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1046 "paddusw %%mm4, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1047 "paddw %%mm3, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1048
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1049 "paddw %%mm2, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1050 "paddw %%mm4, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1051
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1052 "psubusw %%mm3, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1053 "psubusw %%mm2, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1054
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1055 "psubusw %%mm4, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1056 "movq %%mm1, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1057
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1058 "por %%mm5, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1059 "paddusw %%mm7, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1060
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1061 "por %%mm6, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1062 "paddw %%mm7, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1063
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1064 "packssdw %%mm4, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1065 "psubusw %%mm7, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1066
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1067 "movd %%mm4, %%"REG_a" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1068 "or %%"REG_a", %%"REG_a" \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1069 "jnz 2f \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1070 //movq [edi+"DCTSIZE_S"*3*2], mm1
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1071 //movq [edi+"DCTSIZE_S"*5*2], mm5
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1072 //movq [edi+"DCTSIZE_S"*1*2], mm0
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1073 //movq [edi+"DCTSIZE_S"*7*2], mm6
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1074 // t4 t5 - - - t6 t7 -
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1075 //--- t4 (mm0) may be <>0; mm1, mm5, mm6 == 0
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1076 //Typical numbers: nondc - 19%%, dc - 26%%, zero - 55%%. zero case alone isn't worthwhile
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1077 "movq 0*8+%3, %%mm4 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1078 "movq %%mm0, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1079
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1080 "pmulhw "MANGLE(MM_FIX_0_847759065)", %%mm0 \n\t" //tmp6
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1081 "movq %%mm1, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1082
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1083 "movq "DCTSIZE_S"*0*2(%%"REG_D"), %%mm5 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1084 "movq %%mm2, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1085
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1086 "pmulhw "MANGLE(MM_FIX_0_566454497)", %%mm1 \n\t" //tmp5
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1087 "paddw %%mm4, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1088
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1089 "movq 1*8+%3, %%mm6 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1090 //paddw mm3, MM_2
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1091 "psraw $2, %%mm3 \n\t" //tmp7
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1092
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1093 "pmulhw "MANGLE(MM_FIX_0_198912367)", %%mm2 \n\t" //-tmp4
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1094 "psubw %%mm3, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1095
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1096 "movq "DCTSIZE_S"*1*2(%%"REG_D"), %%mm7 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1097 "paddw %%mm3, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1098
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1099 "movq %%mm4, "DCTSIZE_S"*7*2(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1100 "paddw %%mm6, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1101
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1102 "movq 2*8+%3, %%mm3 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1103 "psubw %%mm0, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1104
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1105 "movq "DCTSIZE_S"*2*2(%%"REG_D"), %%mm4 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1106 "paddw %%mm0, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1107
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1108 "movq %%mm5, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1109 "paddw %%mm3, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1110
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1111 "movq %%mm6, "DCTSIZE_S"*6*2(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1112 "psubw %%mm1, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1113
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1114 "movq "DCTSIZE_S"*5*2(%%"REG_D"), %%mm5 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1115 "paddw %%mm1, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1116
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1117 "movq "DCTSIZE_S"*3*2(%%"REG_D"), %%mm6 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1118 "paddw %%mm3, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1119
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1120 "movq 3*8+%3, %%mm0 \n\t"
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1121 "add $8, %%"REG_S" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1122
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1123 "movq %%mm7, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1124 "paddw %%mm0, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1125
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1126 "movq %%mm4, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1127 "psubw %%mm2, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1128
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1129 "movq "DCTSIZE_S"*4*2(%%"REG_D"), %%mm7 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1130 "paddw %%mm2, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1131
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1132 "movq %%mm5, "DCTSIZE_S"*5*2(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1133 "paddw %%mm0, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1134
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1135 "movq %%mm6, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1136
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1137 "movq %%mm7, "DCTSIZE_S"*4*2(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1138 "add $8, %%"REG_D" \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1139 "jmp 4f \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1140
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1141 "2: \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1142 //--- non DC2
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1143 //psraw mm1, 2 w/o it -> offset. thr1, thr1, thr1 (actually thr1, thr1, thr1-1)
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1144 //psraw mm5, 2
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1145 //psraw mm0, 2
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1146 //psraw mm6, 2
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1147 "movq %%mm5, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1148 "psubw %%mm1, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1149
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1150 "psllw $1, %%mm5 \n\t" //'z10
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1151 "paddw %%mm1, %%mm3 \n\t" //'z13
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1152
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1153 "movq %%mm0, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1154 "psubw %%mm6, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1155
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1156 "movq %%mm5, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1157 "psllw $1, %%mm0 \n\t" //'z12
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1158
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1159 "pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm1 \n\t" //-
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1160 "paddw %%mm0, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1161
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1162 "pmulhw "MANGLE(MM_FIX_1_847759065)", %%mm5 \n\t" //'z5
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1163 "paddw %%mm6, %%mm2 \n\t" //'z11
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1164
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1165 "pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1166 "movq %%mm2, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1167
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1168 //---
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1169 "movq 0*8+%3, %%mm4 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1170 "psubw %%mm3, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1171
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1172 "psllw $1, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1173 "paddw %%mm3, %%mm7 \n\t" //'t7
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1174
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1175 "pmulhw "MANGLE(MM_FIX_1_414213562)", %%mm2 \n\t" //'t11
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1176 "movq %%mm4, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1177 //paddw mm7, MM_2
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1178 "psraw $2, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1179
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1180 "paddw "DCTSIZE_S"*0*2(%%"REG_D"), %%mm4 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1181 "psubw %%mm7, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1182
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1183 "movq 1*8+%3, %%mm3 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1184 "paddw %%mm7, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1185
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1186 "movq %%mm6, "DCTSIZE_S"*7*2(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1187 "paddw %%mm5, %%mm1 \n\t" //'t12
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1188
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1189 "movq %%mm4, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1190 "psubw %%mm7, %%mm1 \n\t" //'t6
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1191
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1192 "movq 2*8+%3, %%mm7 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1193 "psubw %%mm5, %%mm0 \n\t" //'t10
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1194
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1195 "movq 3*8+%3, %%mm6 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1196 "movq %%mm3, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1197
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1198 "paddw "DCTSIZE_S"*1*2(%%"REG_D"), %%mm3 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1199 "psubw %%mm1, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1200
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1201 "psubw %%mm1, %%mm2 \n\t" //'t5
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1202 "paddw %%mm1, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1203
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1204 "movq %%mm5, "DCTSIZE_S"*6*2(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1205 "movq %%mm7, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1206
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1207 "paddw "DCTSIZE_S"*2*2(%%"REG_D"), %%mm7 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1208 "psubw %%mm2, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1209
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1210 "paddw "DCTSIZE_S"*5*2(%%"REG_D"), %%mm4 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1211 "paddw %%mm2, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1212
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1213 "movq %%mm3, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1214 "paddw %%mm2, %%mm0 \n\t" //'t4
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1215
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1216 // 't4 't6 't5 - - - - 't7
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1217 "movq %%mm7, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1218 "movq %%mm6, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1219
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1220 "paddw "DCTSIZE_S"*4*2(%%"REG_D"), %%mm6 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1221 "psubw %%mm0, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1222
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1223 "paddw "DCTSIZE_S"*3*2(%%"REG_D"), %%mm1 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1224 "paddw %%mm0, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1225
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1226 "movq %%mm4, "DCTSIZE_S"*5*2(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1227 "add $8, %%"REG_S" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1228
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1229 "movq %%mm6, "DCTSIZE_S"*4*2(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1230
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1231 "movq %%mm1, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1232 "add $8, %%"REG_D" \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1233
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1234 "4: \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1235 //=part 2 (the same)===========================================================
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1236 "movq "DCTSIZE_S"*0*2(%%"REG_S"), %%mm1 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1237 //
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1238 "movq "DCTSIZE_S"*3*2(%%"REG_S"), %%mm7 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1239 "movq %%mm1, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1240
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1241 "paddw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm1 \n\t" //t0
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1242 "movq %%mm7, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1243
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1244 "paddw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm7 \n\t" //t3
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1245 "movq %%mm1, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1246
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1247 "movq "DCTSIZE_S"*1*2(%%"REG_S"), %%mm6 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1248 "psubw %%mm7, %%mm1 \n\t" //t13
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1249
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1250 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1251 "movq %%mm6, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1252
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1253 "paddw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm6 \n\t" //t1
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1254 "paddw %%mm7, %%mm5 \n\t" //t10
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1255
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1256 "paddw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t2
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1257 "movq %%mm6, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1258
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1259 "paddw %%mm2, %%mm6 \n\t" //t11
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1260 "psubw %%mm2, %%mm7 \n\t" //t12
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1261
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1262 "movq %%mm5, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1263 "paddw %%mm6, %%mm5 \n\t" //d0
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1264 // i0 t13 t12 i3 i1 d0 - d4
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1265 "psubw %%mm6, %%mm2 \n\t" //d4
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1266 "paddw %%mm1, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1267
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1268 "movq 1*8+4*16(%%"REG_d"), %%mm6 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1269 "psllw $2, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1270
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1271 "psubw 1*8+0*16(%%"REG_d"), %%mm5 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1272 "psubw %%mm6, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1273
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1274 "paddusw 1*8+0*16(%%"REG_d"), %%mm5 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1275 "paddusw %%mm6, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1276
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1277 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1278 //
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1279 "paddw 1*8+0*16(%%"REG_d"), %%mm5 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1280 "paddw %%mm6, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1281
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1282 "psubusw 1*8+0*16(%%"REG_d"), %%mm5 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1283 "psubusw %%mm6, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1284
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1285 //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
1286 // 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
1287 //However, typical numbers: nondc - 29%%, dc - 46%%, zero - 25%%. All <> 0 case is very rare.
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1288 "paddw "MANGLE(MM_2)", %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1289 "movq %%mm2, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1290
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1291 "paddw %%mm5, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1292 "psubw %%mm6, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1293
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1294 "movq %%mm1, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1295 "paddw %%mm7, %%mm1 \n\t" //d2
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1296
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1297 "psubw 1*8+2*16(%%"REG_d"), %%mm1 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1298 "psubw %%mm7, %%mm6 \n\t" //d6
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1299
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1300 "movq 1*8+6*16(%%"REG_d"), %%mm7 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1301 "psraw $2, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1302
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1303 "paddusw 1*8+2*16(%%"REG_d"), %%mm1 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1304 "psubw %%mm7, %%mm6 \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1305 // t7 d2 /t11 t4 t6 - d6 /t10
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1306
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1307 "paddw 1*8+2*16(%%"REG_d"), %%mm1 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1308 "paddusw %%mm7, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1309
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1310 "psubusw 1*8+2*16(%%"REG_d"), %%mm1 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1311 "paddw %%mm7, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1312
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1313 "psubw "DCTSIZE_S"*4*2(%%"REG_S"), %%mm3 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1314 "psubusw %%mm7, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1315
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1316 //movq [edi+"DCTSIZE_S"*2*2], mm1
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1317 //movq [edi+"DCTSIZE_S"*6*2], mm6
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1318 "movq %%mm1, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1319 "psraw $2, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1320
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1321 "psubw "DCTSIZE_S"*6*2(%%"REG_S"), %%mm4 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1322 "psubw %%mm6, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1323
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1324 "psubw "DCTSIZE_S"*7*2(%%"REG_S"), %%mm0 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1325 "paddw %%mm7, %%mm6 \n\t" //'t13
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1326
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1327 "psraw $2, %%mm6 \n\t" //paddw mm6, MM_2 !! ---
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1328 "movq %%mm2, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1329
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1330 "pmulhw "MANGLE(MM_FIX_1_414213562_A)", %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1331 "paddw %%mm6, %%mm2 \n\t" //'t0
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1332
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1333 "movq %%mm2, 0*8+%3 \n\t" //!
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1334 "psubw %%mm6, %%mm7 \n\t" //'t3
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1335
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1336 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1337 "psubw %%mm6, %%mm1 \n\t" //'t12
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1338
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1339 "psubw "DCTSIZE_S"*5*2(%%"REG_S"), %%mm2 \n\t" //t5
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1340 "movq %%mm5, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1341
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1342 "movq %%mm7, 3*8+%3 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1343 "paddw %%mm2, %%mm3 \n\t" //t10
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1344
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1345 "paddw %%mm4, %%mm2 \n\t" //t11
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1346 "paddw %%mm0, %%mm4 \n\t" //t12
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1347
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1348 "movq %%mm3, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1349 "psubw %%mm4, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1350
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1351 "psllw $2, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1352 "psllw $2, %%mm7 \n\t" //opt for P6
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1353
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1354 "pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1355 "psllw $2, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1356
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1357 "pmulhw "MANGLE(MM_FIX_0_541196100)", %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1358 "psllw $2, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1359
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1360 "pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1361 "paddw %%mm1, %%mm5 \n\t" //'t1
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1362
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1363 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1364 "psubw %%mm1, %%mm6 \n\t" //'t2
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1365 // t7 't12 't11 t4 t6 - 't13 't10 ---
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1366
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1367 "paddw %%mm3, %%mm7 \n\t" //z2
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1368
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1369 "movq %%mm5, 1*8+%3 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1370 "paddw %%mm3, %%mm4 \n\t" //z4
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1371
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1372 "movq 1*8+3*16(%%"REG_d"), %%mm3 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1373 "movq %%mm0, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1374
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1375 "movq %%mm6, 2*8+%3 \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1376 "psubw %%mm2, %%mm1 \n\t" //z13
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1377
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1378 //===
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1379 "paddw %%mm2, %%mm0 \n\t" //z11
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1380 "movq %%mm1, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1381
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1382 "movq 1*8+5*16(%%"REG_d"), %%mm2 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1383 "psubw %%mm7, %%mm1 \n\t" //d3
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1384
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1385 "paddw %%mm7, %%mm5 \n\t" //d5
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1386 "psubw %%mm3, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1387
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1388 "movq 1*8+1*16(%%"REG_d"), %%mm7 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1389 "psubw %%mm2, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1390
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1391 "movq %%mm0, %%mm6 \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1392 "paddw %%mm4, %%mm0 \n\t" //d1
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1393
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1394 "paddusw %%mm3, %%mm1 \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1395 "psubw %%mm4, %%mm6 \n\t" //d7
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1396
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1397 // d1 d3 - - - d5 d7 -
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1398 "movq 1*8+7*16(%%"REG_d"), %%mm4 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1399 "psubw %%mm7, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1400
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1401 "psubw %%mm4, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1402 "paddusw %%mm2, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1403
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1404 "paddusw %%mm4, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1405 "paddw %%mm3, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1406
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1407 "paddw %%mm2, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1408 "paddw %%mm4, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1409
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1410 "psubusw %%mm3, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1411 "psubusw %%mm2, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1412
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1413 "psubusw %%mm4, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1414 "movq %%mm1, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1415
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1416 "por %%mm5, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1417 "paddusw %%mm7, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1418
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1419 "por %%mm6, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1420 "paddw %%mm7, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1421
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1422 "packssdw %%mm4, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1423 "psubusw %%mm7, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1424
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1425 "movd %%mm4, %%"REG_a" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1426 "or %%"REG_a", %%"REG_a" \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1427 "jnz 3f \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1428 //movq [edi+"DCTSIZE_S"*3*2], mm1
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1429 //movq [edi+"DCTSIZE_S"*5*2], mm5
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1430 //movq [edi+"DCTSIZE_S"*1*2], mm0
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1431 //movq [edi+"DCTSIZE_S"*7*2], mm6
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1432 // t4 t5 - - - t6 t7 -
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1433 //--- t4 (mm0) may be <>0; mm1, mm5, mm6 == 0
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1434 //Typical numbers: nondc - 19%%, dc - 26%%, zero - 55%%. zero case alone isn't worthwhile
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1435 "movq 0*8+%3, %%mm4 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1436 "movq %%mm0, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1437
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1438 "pmulhw "MANGLE(MM_FIX_0_847759065)", %%mm0 \n\t" //tmp6
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1439 "movq %%mm1, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1440
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1441 "movq "DCTSIZE_S"*0*2(%%"REG_D"), %%mm5 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1442 "movq %%mm2, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1443
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1444 "pmulhw "MANGLE(MM_FIX_0_566454497)", %%mm1 \n\t" //tmp5
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1445 "paddw %%mm4, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1446
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1447 "movq 1*8+%3, %%mm6 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1448 //paddw mm3, MM_2
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1449 "psraw $2, %%mm3 \n\t" //tmp7
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1450
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1451 "pmulhw "MANGLE(MM_FIX_0_198912367)", %%mm2 \n\t" //-tmp4
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1452 "psubw %%mm3, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1453
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1454 "movq "DCTSIZE_S"*1*2(%%"REG_D"), %%mm7 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1455 "paddw %%mm3, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1456
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1457 "movq %%mm4, "DCTSIZE_S"*7*2(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1458 "paddw %%mm6, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1459
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1460 "movq 2*8+%3, %%mm3 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1461 "psubw %%mm0, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1462
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1463 "movq "DCTSIZE_S"*2*2(%%"REG_D"), %%mm4 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1464 "paddw %%mm0, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1465
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1466 "movq %%mm5, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1467 "paddw %%mm3, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1468
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1469 "movq %%mm6, "DCTSIZE_S"*6*2(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1470 "psubw %%mm1, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1471
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1472 "movq "DCTSIZE_S"*5*2(%%"REG_D"), %%mm5 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1473 "paddw %%mm1, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1474
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1475 "movq "DCTSIZE_S"*3*2(%%"REG_D"), %%mm6 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1476 "paddw %%mm3, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1477
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1478 "movq 3*8+%3, %%mm0 \n\t"
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1479 "add $24, %%"REG_S" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1480
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1481 "movq %%mm7, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1482 "paddw %%mm0, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1483
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1484 "movq %%mm4, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1485 "psubw %%mm2, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1486
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1487 "movq "DCTSIZE_S"*4*2(%%"REG_D"), %%mm7 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1488 "paddw %%mm2, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1489
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1490 "movq %%mm5, "DCTSIZE_S"*5*2(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1491 "paddw %%mm0, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1492
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1493 "movq %%mm6, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1494
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1495 "movq %%mm7, "DCTSIZE_S"*4*2(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1496 "add $24, %%"REG_D" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1497 "sub $2, %%"REG_c" \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1498 "jnz 1b \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1499 "jmp 5f \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1500
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1501 "3: \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1502 //--- non DC2
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1503 //psraw mm1, 2 w/o it -> offset. thr1, thr1, thr1 (actually thr1, thr1, thr1-1)
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1504 //psraw mm5, 2
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1505 //psraw mm0, 2
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1506 //psraw mm6, 2
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1507 "movq %%mm5, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1508 "psubw %%mm1, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1509
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1510 "psllw $1, %%mm5 \n\t" //'z10
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1511 "paddw %%mm1, %%mm3 \n\t" //'z13
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1512
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1513 "movq %%mm0, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1514 "psubw %%mm6, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1515
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1516 "movq %%mm5, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1517 "psllw $1, %%mm0 \n\t" //'z12
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1518
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1519 "pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm1 \n\t" //-
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1520 "paddw %%mm0, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1521
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1522 "pmulhw "MANGLE(MM_FIX_1_847759065)", %%mm5 \n\t" //'z5
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1523 "paddw %%mm6, %%mm2 \n\t" //'z11
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1524
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1525 "pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1526 "movq %%mm2, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1527
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1528 //---
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1529 "movq 0*8+%3, %%mm4 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1530 "psubw %%mm3, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1531
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1532 "psllw $1, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1533 "paddw %%mm3, %%mm7 \n\t" //'t7
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1534
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1535 "pmulhw "MANGLE(MM_FIX_1_414213562)", %%mm2 \n\t" //'t11
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1536 "movq %%mm4, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1537 //paddw mm7, MM_2
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1538 "psraw $2, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1539
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1540 "paddw "DCTSIZE_S"*0*2(%%"REG_D"), %%mm4 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1541 "psubw %%mm7, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1542
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1543 "movq 1*8+%3, %%mm3 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1544 "paddw %%mm7, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1545
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1546 "movq %%mm6, "DCTSIZE_S"*7*2(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1547 "paddw %%mm5, %%mm1 \n\t" //'t12
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1548
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1549 "movq %%mm4, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1550 "psubw %%mm7, %%mm1 \n\t" //'t6
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1551
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1552 "movq 2*8+%3, %%mm7 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1553 "psubw %%mm5, %%mm0 \n\t" //'t10
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1554
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1555 "movq 3*8+%3, %%mm6 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1556 "movq %%mm3, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1557
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1558 "paddw "DCTSIZE_S"*1*2(%%"REG_D"), %%mm3 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1559 "psubw %%mm1, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1560
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1561 "psubw %%mm1, %%mm2 \n\t" //'t5
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1562 "paddw %%mm1, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1563
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1564 "movq %%mm5, "DCTSIZE_S"*6*2(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1565 "movq %%mm7, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1566
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1567 "paddw "DCTSIZE_S"*2*2(%%"REG_D"), %%mm7 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1568 "psubw %%mm2, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1569
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1570 "paddw "DCTSIZE_S"*5*2(%%"REG_D"), %%mm4 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1571 "paddw %%mm2, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1572
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1573 "movq %%mm3, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1574 "paddw %%mm2, %%mm0 \n\t" //'t4
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1575
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1576 // 't4 't6 't5 - - - - 't7
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1577 "movq %%mm7, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1578 "movq %%mm6, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1579
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1580 "paddw "DCTSIZE_S"*4*2(%%"REG_D"), %%mm6 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1581 "psubw %%mm0, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1582
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1583 "paddw "DCTSIZE_S"*3*2(%%"REG_D"), %%mm1 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1584 "paddw %%mm0, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1585
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1586 "movq %%mm4, "DCTSIZE_S"*5*2(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1587 "add $24, %%"REG_S" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1588
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1589 "movq %%mm6, "DCTSIZE_S"*4*2(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1590
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1591 "movq %%mm1, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1592 "add $24, %%"REG_D" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1593 "sub $2, %%"REG_c" \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1594 "jnz 1b \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1595 "5: \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1596
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1597 : "+S"(data), "+D"(output), "+c"(cnt), "=o"(temps)
15632
e813a3e431a8 move unchanged registers back to input spec
henry
parents: 15631
diff changeset
1598 : "d"(thr_adr)
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1599 : "%"REG_a
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1600 );
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1601 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1602
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1603 #endif // HAVE_MMX
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1604
28290
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_...
reimar
parents: 27754
diff changeset
1605 #if !HAVE_MMX
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1606
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1607 static void row_idct_c(DCTELEM* workspace,
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1608 int16_t* output_adr, int output_stride, int cnt)
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1609 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1610 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
1611 int_simd16_t tmp10, tmp11, tmp12, tmp13;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1612 int_simd16_t z5, z10, z11, z12, z13;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1613 int16_t* outptr;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1614 DCTELEM* wsptr;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1615
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1616 cnt*=4;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1617 wsptr = workspace;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1618 outptr = output_adr;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1619 for (; cnt > 0; cnt--) {
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1620 // Even part
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1621 //Simd version reads 4x4 block and transposes it
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1622 tmp10 = ( wsptr[2] + wsptr[3]);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1623 tmp11 = ( wsptr[2] - wsptr[3]);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1624
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1625 tmp13 = ( wsptr[0] + wsptr[1]);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1626 tmp12 = (MULTIPLY16H( wsptr[0] - wsptr[1], FIX_1_414213562_A)<<2) - tmp13;//this shift order to avoid overflow
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1627
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1628 tmp0 = tmp10 + tmp13; //->temps
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1629 tmp3 = tmp10 - tmp13; //->temps
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1630 tmp1 = tmp11 + tmp12;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1631 tmp2 = tmp11 - tmp12;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1632
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1633 // Odd part
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1634 //Also transpose, with previous:
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1635 // ---- ---- ||||
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1636 // ---- ---- idct ||||
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1637 // ---- ---- ---> ||||
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1638 // ---- ---- ||||
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1639 z13 = wsptr[4] + wsptr[5];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1640 z10 = wsptr[4] - wsptr[5];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1641 z11 = wsptr[6] + wsptr[7];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1642 z12 = wsptr[6] - wsptr[7];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1643
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1644 tmp7 = z11 + z13;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1645 tmp11 = MULTIPLY16H(z11 - z13, FIX_1_414213562);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1646
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1647 z5 = MULTIPLY16H(z10 + z12, FIX_1_847759065);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1648 tmp10 = MULTIPLY16H(z12, FIX_1_082392200) - z5;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1649 tmp12 = MULTIPLY16H(z10, FIX_2_613125930) + z5; // - FIX_
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1650
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1651 tmp6 = (tmp12<<3) - tmp7;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1652 tmp5 = (tmp11<<3) - tmp6;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1653 tmp4 = (tmp10<<3) + tmp5;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1654
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1655 // Final output stage: descale and write column
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1656 outptr[0*output_stride]+= DESCALE(tmp0 + tmp7, 3);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1657 outptr[1*output_stride]+= DESCALE(tmp1 + tmp6, 3);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1658 outptr[2*output_stride]+= DESCALE(tmp2 + tmp5, 3);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1659 outptr[3*output_stride]+= DESCALE(tmp3 - tmp4, 3);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1660 outptr[4*output_stride]+= DESCALE(tmp3 + tmp4, 3);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1661 outptr[5*output_stride]+= DESCALE(tmp2 - tmp5, 3);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1662 outptr[6*output_stride]+= DESCALE(tmp1 - tmp6, 3); //no += ?
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1663 outptr[7*output_stride]+= DESCALE(tmp0 - tmp7, 3); //no += ?
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1664 outptr++;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1665
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1666 wsptr += DCTSIZE; // advance pointer to next row
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1667 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1668 }
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 #else /* HAVE_MMX */
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1671
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1672 static void row_idct_mmx (DCTELEM* workspace,
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1673 int16_t* output_adr, int output_stride, int cnt)
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1674 {
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1675 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
1676 __asm__ volatile(
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1677 "lea (%%"REG_a",%%"REG_a",2), %%"REG_d" \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1678 "1: \n\t"
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1679 "movq "DCTSIZE_S"*0*2(%%"REG_S"), %%mm0 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1680 //
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1681
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1682 "movq "DCTSIZE_S"*1*2(%%"REG_S"), %%mm1 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1683 "movq %%mm0, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1684
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1685 "movq "DCTSIZE_S"*2*2(%%"REG_S"), %%mm2 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1686 "punpcklwd %%mm1, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1687
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1688 "movq "DCTSIZE_S"*3*2(%%"REG_S"), %%mm3 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1689 "punpckhwd %%mm1, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1690
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1691 //transpose 4x4
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1692 "movq %%mm2, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1693 "punpcklwd %%mm3, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1694
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1695 "movq %%mm0, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1696 "punpckldq %%mm2, %%mm0 \n\t" //0
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1697
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1698 "punpckhdq %%mm2, %%mm6 \n\t" //1
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1699 "movq %%mm0, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1700
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1701 "punpckhwd %%mm3, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1702 "psubw %%mm6, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1703
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1704 "pmulhw "MANGLE(MM_FIX_1_414213562_A)", %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1705 "movq %%mm4, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1706
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1707 "punpckldq %%mm7, %%mm4 \n\t" //2
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1708 "paddw %%mm6, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1709
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1710 "punpckhdq %%mm7, %%mm2 \n\t" //3
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1711 "movq %%mm4, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1712
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1713 "psllw $2, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1714 "paddw %%mm2, %%mm4 \n\t" //t10
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1715
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1716 "movq "DCTSIZE_S"*0*2+"DCTSIZE_S"(%%"REG_S"), %%mm3 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1717 "psubw %%mm2, %%mm1 \n\t" //t11
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1718
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1719 "movq "DCTSIZE_S"*1*2+"DCTSIZE_S"(%%"REG_S"), %%mm2 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1720 "psubw %%mm5, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1721
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1722 "movq %%mm4, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1723 "paddw %%mm5, %%mm4 \n\t" //t0
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1724
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1725 "psubw %%mm5, %%mm6 \n\t" //t3
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1726 "movq %%mm1, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1727
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1728 "movq "DCTSIZE_S"*2*2+"DCTSIZE_S"(%%"REG_S"), %%mm5 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1729 "paddw %%mm0, %%mm1 \n\t" //t1
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1730
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1731 "movq %%mm4, 0*8+%3 \n\t" //t0
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1732 "movq %%mm3, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1733
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1734 "movq %%mm6, 1*8+%3 \n\t" //t3
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1735 "punpcklwd %%mm2, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1736
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1737 //transpose 4x4
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1738 "movq "DCTSIZE_S"*3*2+"DCTSIZE_S"(%%"REG_S"), %%mm6 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1739 "punpckhwd %%mm2, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1740
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1741 "movq %%mm5, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1742 "punpcklwd %%mm6, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1743
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1744 "psubw %%mm0, %%mm7 \n\t" //t2
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1745 "punpckhwd %%mm6, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1746
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1747 "movq %%mm3, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1748 "punpckldq %%mm5, %%mm3 \n\t" //4
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1749
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1750 "punpckhdq %%mm5, %%mm0 \n\t" //5
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1751 "movq %%mm4, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1752
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1753 //
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1754 "movq %%mm3, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1755 "punpckldq %%mm2, %%mm4 \n\t" //6
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1756
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1757 "psubw %%mm0, %%mm3 \n\t" //z10
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1758 "punpckhdq %%mm2, %%mm5 \n\t" //7
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1759
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1760 "paddw %%mm0, %%mm6 \n\t" //z13
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1761 "movq %%mm4, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1762
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1763 "movq %%mm3, %%mm0 \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1764 "psubw %%mm5, %%mm4 \n\t" //z12
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1765
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1766 "pmulhw "MANGLE(MM_FIX_2_613125930)", %%mm0 \n\t" //-
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1767 "paddw %%mm4, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1768
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1769 "pmulhw "MANGLE(MM_FIX_1_847759065)", %%mm3 \n\t" //z5
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1770 "paddw %%mm5, %%mm2 \n\t" //z11 >
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1771
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1772 "pmulhw "MANGLE(MM_FIX_1_082392200)", %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1773 "movq %%mm2, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1774
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1775 "psubw %%mm6, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1776 "paddw %%mm6, %%mm5 \n\t" //t7
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1777
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1778 "pmulhw "MANGLE(MM_FIX_1_414213562)", %%mm2 \n\t" //t11
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1779 "paddw %%mm3, %%mm0 \n\t" //t12
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1780
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1781 "psllw $3, %%mm0 \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1782 "psubw %%mm3, %%mm4 \n\t" //t10
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1783
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1784 "movq 0*8+%3, %%mm6 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1785 "movq %%mm1, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1786
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1787 "psllw $3, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1788 "psubw %%mm5, %%mm0 \n\t" //t6
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1789
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1790 "psllw $3, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1791 "paddw %%mm0, %%mm1 \n\t" //d1
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1792
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1793 "psubw %%mm0, %%mm2 \n\t" //t5
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1794 "psubw %%mm0, %%mm3 \n\t" //d6
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1795
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1796 "paddw %%mm2, %%mm4 \n\t" //t4
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1797 "movq %%mm7, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1798
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1799 "paddw %%mm2, %%mm7 \n\t" //d2
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1800 "psubw %%mm2, %%mm0 \n\t" //d5
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1801
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1802 "movq "MANGLE(MM_DESCALE_RND)", %%mm2 \n\t" //4
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1803 "psubw %%mm5, %%mm6 \n\t" //d7
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1804
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1805 "paddw 0*8+%3, %%mm5 \n\t" //d0
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1806 "paddw %%mm2, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1807
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1808 "paddw %%mm2, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1809 "psraw $3, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1810
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1811 "paddw %%mm2, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1812 "psraw $3, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1813
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1814 "paddw (%%"REG_D"), %%mm5 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1815 "psraw $3, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1816
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1817 "paddw (%%"REG_D",%%"REG_a",), %%mm1 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1818 "paddw %%mm2, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1819
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1820 "paddw (%%"REG_D",%%"REG_a",2), %%mm7 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1821 "paddw %%mm2, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1822
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1823 "movq %%mm5, (%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1824 "paddw %%mm2, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1825
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1826 "movq %%mm1, (%%"REG_D",%%"REG_a",) \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1827 "psraw $3, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1828
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1829 "movq %%mm7, (%%"REG_D",%%"REG_a",2) \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1830 "add %%"REG_d", %%"REG_D" \n\t" //3*ls
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1831
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1832 "movq 1*8+%3, %%mm5 \n\t" //t3
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1833 "psraw $3, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1834
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1835 "paddw (%%"REG_D",%%"REG_a",2), %%mm0 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1836 "psubw %%mm4, %%mm5 \n\t" //d3
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1837
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1838 "paddw (%%"REG_D",%%"REG_d",), %%mm3 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1839 "psraw $3, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1840
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1841 "paddw 1*8+%3, %%mm4 \n\t" //d4
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1842 "paddw %%mm2, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1843
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1844 "paddw (%%"REG_D",%%"REG_a",4), %%mm6 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1845 "paddw %%mm2, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1846
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1847 "movq %%mm0, (%%"REG_D",%%"REG_a",2) \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1848 "psraw $3, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1849
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1850 "paddw (%%"REG_D"), %%mm5 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1851 "psraw $3, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1852
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1853 "paddw (%%"REG_D",%%"REG_a",), %%mm4 \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1854 "add $"DCTSIZE_S"*2*4, %%"REG_S" \n\t" //4 rows
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1855
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1856 "movq %%mm3, (%%"REG_D",%%"REG_d",) \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1857 "movq %%mm6, (%%"REG_D",%%"REG_a",4) \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1858 "movq %%mm5, (%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1859 "movq %%mm4, (%%"REG_D",%%"REG_a",) \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1860
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1861 "sub %%"REG_d", %%"REG_D" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1862 "add $8, %%"REG_D" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1863 "dec %%"REG_c" \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1864 "jnz 1b \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1865
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1866 : "+S"(workspace), "+D"(output_adr), "+c"(cnt), "=o"(temps)
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1867 : "a"(output_stride*sizeof(short))
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1868 : "%"REG_d
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1869 );
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1870 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1871
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1872 #endif // HAVE_MMX
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1873
28290
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_...
reimar
parents: 27754
diff changeset
1874 #if !HAVE_MMX
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1875
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1876 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
1877 {
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1878 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
1879 int_simd16_t tmp10, tmp11, tmp12, tmp13;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1880 int_simd16_t z1, z2, z3, z4, z5, z11, z13;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1881 DCTELEM *dataptr;
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1882
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1883 cnt*=4;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1884 // Pass 1: process rows.
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1885
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1886 dataptr = data;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1887 for (; cnt > 0; cnt--) {
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1888 tmp0 = pixels[line_size*0] + pixels[line_size*7];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1889 tmp7 = pixels[line_size*0] - pixels[line_size*7];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1890 tmp1 = pixels[line_size*1] + pixels[line_size*6];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1891 tmp6 = pixels[line_size*1] - pixels[line_size*6];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1892 tmp2 = pixels[line_size*2] + pixels[line_size*5];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1893 tmp5 = pixels[line_size*2] - pixels[line_size*5];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1894 tmp3 = pixels[line_size*3] + pixels[line_size*4];
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1895 tmp4 = pixels[line_size*3] - pixels[line_size*4];
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1896
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1897 // Even part
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1898
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1899 tmp10 = tmp0 + tmp3;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1900 tmp13 = tmp0 - tmp3;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1901 tmp11 = tmp1 + tmp2;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1902 tmp12 = tmp1 - tmp2;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1903 //Even columns are written first, this leads to different order of columns
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1904 //in column_fidct(), but they are processed independently, so all ok.
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1905 //Later in the row_idct() columns readed at the same order.
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1906 dataptr[2] = tmp10 + tmp11;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1907 dataptr[3] = tmp10 - tmp11;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1908
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1909 z1 = MULTIPLY16H((tmp12 + tmp13)<<2, FIX_0_707106781);
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1910 dataptr[0] = tmp13 + z1;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1911 dataptr[1] = tmp13 - z1;
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1912
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1913 // Odd part
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1914
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1915 tmp10 = (tmp4 + tmp5) <<2;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1916 tmp11 = (tmp5 + tmp6) <<2;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1917 tmp12 = (tmp6 + tmp7) <<2;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1918
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1919 z5 = MULTIPLY16H(tmp10 - tmp12, FIX_0_382683433);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1920 z2 = MULTIPLY16H(tmp10, FIX_0_541196100) + z5;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1921 z4 = MULTIPLY16H(tmp12, FIX_1_306562965) + z5;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1922 z3 = MULTIPLY16H(tmp11, FIX_0_707106781);
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1923
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1924 z11 = tmp7 + z3;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1925 z13 = tmp7 - z3;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1926
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1927 dataptr[4] = z13 + z2;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1928 dataptr[5] = z13 - z2;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1929 dataptr[6] = z11 + z4;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1930 dataptr[7] = z11 - z4;
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1931
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1932 pixels++; // advance pointer to next column
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1933 dataptr += DCTSIZE;
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1934 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1935 }
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 #else /* HAVE_MMX */
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 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
1940 {
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1941 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
1942 __asm__ volatile(
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1943 "lea (%%"REG_a",%%"REG_a",2), %%"REG_d" \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1944 "6: \n\t"
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1945 "movd (%%"REG_S"), %%mm0 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1946 "pxor %%mm7, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1947
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1948 "movd (%%"REG_S",%%"REG_a",), %%mm1 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1949 "punpcklbw %%mm7, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1950
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1951 "movd (%%"REG_S",%%"REG_a",2), %%mm2 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1952 "punpcklbw %%mm7, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1953
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1954 "punpcklbw %%mm7, %%mm2 \n\t"
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1955 "add %%"REG_d", %%"REG_S" \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1956
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1957 "movq %%mm0, %%mm5 \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
1958 //
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1959
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1960 "movd (%%"REG_S",%%"REG_a",4), %%mm3 \n\t" //7 ;prefetch!
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1961 "movq %%mm1, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1962
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1963 "movd (%%"REG_S",%%"REG_d",), %%mm4 \n\t" //6
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1964 "punpcklbw %%mm7, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1965
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1966 "psubw %%mm3, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1967 "punpcklbw %%mm7, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1968
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1969 "paddw %%mm3, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1970 "psubw %%mm4, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1971
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1972 "movd (%%"REG_S",%%"REG_a",2), %%mm3 \n\t" //5
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1973 "paddw %%mm4, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1974
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1975 "movq %%mm5, 0*8+%3 \n\t" //t7
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1976 "punpcklbw %%mm7, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1977
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
1978 "movq %%mm6, 1*8+%3 \n\t" //t6
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1979 "movq %%mm2, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1980
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1981 "movd (%%"REG_S"), %%mm5 \n\t" //3
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1982 "paddw %%mm3, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1983
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
1984 "movd (%%"REG_S",%%"REG_a",), %%mm6 \n\t" //4
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1985 "punpcklbw %%mm7, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1986
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1987 "psubw %%mm3, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1988 "punpcklbw %%mm7, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1989
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1990 "movq %%mm5, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1991 "paddw %%mm6, %%mm5 \n\t" //t3
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1992
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1993 "psubw %%mm6, %%mm3 \n\t" //t4 ; t0 t1 t2 t4 t5 t3 - -
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1994 "movq %%mm0, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1995
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1996 "movq %%mm1, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1997 "psubw %%mm5, %%mm0 \n\t" //t13
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1998
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
1999 "psubw %%mm2, %%mm1 \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
2000 "paddw %%mm2, %%mm7 \n\t" //t11
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2001
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2002 "paddw %%mm0, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2003 "movq %%mm7, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2004
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2005 "psllw $2, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2006 "paddw %%mm5, %%mm6 \n\t" //t10
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2007
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2008 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2009 "paddw %%mm6, %%mm7 \n\t" //d2
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2010
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2011 "psubw %%mm2, %%mm6 \n\t" //d3
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2012 "movq %%mm0, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2013
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2014 //transpose 4x4
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2015 "movq %%mm7, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2016 "punpcklwd %%mm6, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2017
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2018 "paddw %%mm1, %%mm0 \n\t" //d0
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2019 "punpckhwd %%mm6, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2020
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
2021 "psubw %%mm1, %%mm5 \n\t" //d1
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2022 "movq %%mm0, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2023
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
2024 "movq 1*8+%3, %%mm1 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2025 "punpcklwd %%mm5, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2026
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2027 "punpckhwd %%mm5, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2028 "movq %%mm0, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2029
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2030 "punpckldq %%mm7, %%mm0 \n\t" //0
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2031 "paddw %%mm4, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2032
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2033 "punpckhdq %%mm7, %%mm5 \n\t" //1
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2034 "movq %%mm6, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2035
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
2036 "movq %%mm0, "DCTSIZE_S"*0*2(%%"REG_D") \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
2037 "punpckldq %%mm2, %%mm6 \n\t" //2
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2038
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
2039 "movq %%mm5, "DCTSIZE_S"*1*2(%%"REG_D") \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
2040 "punpckhdq %%mm2, %%mm7 \n\t" //3
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2041
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
2042 "movq %%mm6, "DCTSIZE_S"*2*2(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2043 "paddw %%mm1, %%mm4 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2044
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
2045 "movq %%mm7, "DCTSIZE_S"*3*2(%%"REG_D") \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
2046 "psllw $2, %%mm3 \n\t" //t10
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2047
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
2048 "movq 0*8+%3, %%mm2 \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2049 "psllw $2, %%mm4 \n\t" //t11
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2050
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2051 "pmulhw "MANGLE(MM_FIX_0_707106781)", %%mm4 \n\t" //z3
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2052 "paddw %%mm2, %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2053
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2054 "psllw $2, %%mm1 \n\t" //t12
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2055 "movq %%mm3, %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2056
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2057 "pmulhw "MANGLE(MM_FIX_0_541196100)", %%mm0 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2058 "psubw %%mm1, %%mm3 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2059
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2060 "pmulhw "MANGLE(MM_FIX_0_382683433)", %%mm3 \n\t" //z5
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2061 "movq %%mm2, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2062
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2063 "pmulhw "MANGLE(MM_FIX_1_306562965)", %%mm1 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2064 "psubw %%mm4, %%mm2 \n\t" //z13
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2065
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2066 "paddw %%mm4, %%mm5 \n\t" //z11
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2067 "movq %%mm2, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2068
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2069 "paddw %%mm3, %%mm0 \n\t" //z2
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2070 "movq %%mm5, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2071
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2072 "paddw %%mm0, %%mm2 \n\t" //d4
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
2073 "psubw %%mm0, %%mm6 \n\t" //d5
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2074
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2075 "movq %%mm2, %%mm4 \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
2076 "paddw %%mm3, %%mm1 \n\t" //z4
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2077
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2078 //transpose 4x4
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2079 "punpcklwd %%mm6, %%mm2 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2080 "paddw %%mm1, %%mm5 \n\t" //d6
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2081
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2082 "punpckhwd %%mm6, %%mm4 \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
2083 "psubw %%mm1, %%mm7 \n\t" //d7
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2084
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2085 "movq %%mm5, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2086 "punpcklwd %%mm7, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2087
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2088 "punpckhwd %%mm7, %%mm6 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2089 "movq %%mm2, %%mm7 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2090
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2091 "punpckldq %%mm5, %%mm2 \n\t" //4
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
2092 "sub %%"REG_d", %%"REG_S" \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2093
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2094 "punpckhdq %%mm5, %%mm7 \n\t" //5
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2095 "movq %%mm4, %%mm5 \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2096
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
2097 "movq %%mm2, "DCTSIZE_S"*0*2+"DCTSIZE_S"(%%"REG_D") \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2098 "punpckldq %%mm6, %%mm4 \n\t" //6
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2099
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
2100 "movq %%mm7, "DCTSIZE_S"*1*2+"DCTSIZE_S"(%%"REG_D") \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
2101 "punpckhdq %%mm6, %%mm5 \n\t" //7
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2102
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
2103 "movq %%mm4, "DCTSIZE_S"*2*2+"DCTSIZE_S"(%%"REG_D") \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
2104 "add $4, %%"REG_S" \n\t"
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
2105
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
2106 "movq %%mm5, "DCTSIZE_S"*3*2+"DCTSIZE_S"(%%"REG_D") \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 29087
diff changeset
2107 "add $"DCTSIZE_S"*2*4, %%"REG_D" \n\t" //4 rows
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
2108 "dec %%"REG_c" \n\t"
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2109 "jnz 6b \n\t"
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2110
26052
ce480034f391 Do not use a global temps variable, this is ugly and does not compile with ICC.
reimar
parents: 26050
diff changeset
2111 : "+S"(pixels), "+D"(data), "+c"(cnt), "=o"(temps)
15632
e813a3e431a8 move unchanged registers back to input spec
henry
parents: 15631
diff changeset
2112 : "a"(line_size)
15634
7eddcf69a5fd x86-64 fixes by Reimar
henry
parents: 15633
diff changeset
2113 : "%"REG_d);
15631
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2114 }
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2115
d5a95e6f5f07 faster spp filter by Nikolaj Poroshin <porosh3 at psu ru>
henry
parents:
diff changeset
2116 #endif // HAVE_MMX