annotate dsputil.c @ 493:873b9075d853 libavcodec

move unaligned access macros to dsputil.h - added unaligned 32 bit store
author bellard
date Tue, 11 Jun 2002 13:43:27 +0000
parents a7cbee351b55
children 0cb7ab35ccb2
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1 /*
986e461dc072 Initial revision
glantau
parents:
diff changeset
2 * DSP utils
429
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
3 * Copyright (c) 2000, 2001 Fabrice Bellard.
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
4 *
429
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
5 * This library is free software; you can redistribute it and/or
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
6 * modify it under the terms of the GNU Lesser General Public
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
7 * License as published by the Free Software Foundation; either
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
8 * version 2 of the License, or (at your option) any later version.
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
9 *
429
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
10 * This library is distributed in the hope that it will be useful,
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
429
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
13 * Lesser General Public License for more details.
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
14 *
429
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
15 * You should have received a copy of the GNU Lesser General Public
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
16 * License along with this library; if not, write to the Free Software
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
256
4c1cec7c3c7c q-pel mc fixed
michaelni
parents: 255
diff changeset
18 *
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
19 * gmc & q-pel & 32/64 bit based MC by Michael Niedermayer <michaelni@gmx.at>
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
20 */
986e461dc072 Initial revision
glantau
parents:
diff changeset
21 #include "avcodec.h"
986e461dc072 Initial revision
glantau
parents:
diff changeset
22 #include "dsputil.h"
174
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
23 #include "simple_idct.h"
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
24
19
82d4c9be9873 MMX/MMXEXT iDCT support, using external functions currently defined in libmpeg2
arpi_esp
parents: 6
diff changeset
25 void (*ff_idct)(DCTELEM *block);
480
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
26 void (*ff_idct_put)(UINT8 *dest, int line_size, DCTELEM *block);
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
27 void (*ff_idct_add)(UINT8 *dest, int line_size, DCTELEM *block);
474
11dbd00682fc avoid name clash with libjpeg - added missing externs
bellard
parents: 468
diff changeset
28 void (*av_fdct)(DCTELEM *block);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
29 void (*get_pixels)(DCTELEM *block, const UINT8 *pixels, int line_size);
324
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
30 void (*diff_pixels)(DCTELEM *block, const UINT8 *s1, const UINT8 *s2, int stride);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
31 void (*put_pixels_clamped)(const DCTELEM *block, UINT8 *pixels, int line_size);
986e461dc072 Initial revision
glantau
parents:
diff changeset
32 void (*add_pixels_clamped)(const DCTELEM *block, UINT8 *pixels, int line_size);
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
33 void (*gmc1)(UINT8 *dst, UINT8 *src, int srcStride, int h, int x16, int y16, int rounder);
296
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
34 void (*clear_blocks)(DCTELEM *blocks);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
35
986e461dc072 Initial revision
glantau
parents:
diff changeset
36 op_pixels_abs_func pix_abs16x16;
986e461dc072 Initial revision
glantau
parents:
diff changeset
37 op_pixels_abs_func pix_abs16x16_x2;
986e461dc072 Initial revision
glantau
parents:
diff changeset
38 op_pixels_abs_func pix_abs16x16_y2;
986e461dc072 Initial revision
glantau
parents:
diff changeset
39 op_pixels_abs_func pix_abs16x16_xy2;
986e461dc072 Initial revision
glantau
parents:
diff changeset
40
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
41 op_pixels_abs_func pix_abs8x8;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
42 op_pixels_abs_func pix_abs8x8_x2;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
43 op_pixels_abs_func pix_abs8x8_y2;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
44 op_pixels_abs_func pix_abs8x8_xy2;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
45
50
4ea4c10d03d8 export for imgconvert usage
glantau
parents: 34
diff changeset
46 UINT8 cropTbl[256 + 2 * MAX_NEG_CROP];
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
47 UINT32 squareTbl[512];
986e461dc072 Initial revision
glantau
parents:
diff changeset
48
435
9247ad420889 * compatibilized declaration with its original definition
kabi
parents: 429
diff changeset
49 extern INT16 default_intra_matrix[64];
9247ad420889 * compatibilized declaration with its original definition
kabi
parents: 429
diff changeset
50 extern INT16 default_non_intra_matrix[64];
9247ad420889 * compatibilized declaration with its original definition
kabi
parents: 429
diff changeset
51 extern INT16 ff_mpeg4_default_intra_matrix[64];
9247ad420889 * compatibilized declaration with its original definition
kabi
parents: 429
diff changeset
52 extern INT16 ff_mpeg4_default_non_intra_matrix[64];
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
53
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
54 UINT8 zigzag_direct[64] = {
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
55 0, 1, 8, 16, 9, 2, 3, 10,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
56 17, 24, 32, 25, 18, 11, 4, 5,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
57 12, 19, 26, 33, 40, 48, 41, 34,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
58 27, 20, 13, 6, 7, 14, 21, 28,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
59 35, 42, 49, 56, 57, 50, 43, 36,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
60 29, 22, 15, 23, 30, 37, 44, 51,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
61 58, 59, 52, 45, 38, 31, 39, 46,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
62 53, 60, 61, 54, 47, 55, 62, 63
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
63 };
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
64
220
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
65 /* not permutated inverse zigzag_direct + 1 for MMX quantizer */
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
66 UINT16 __align8 inv_zigzag_direct16[64];
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
67
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
68 /* not permutated zigzag_direct for MMX quantizer */
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
69 UINT8 zigzag_direct_noperm[64];
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
70
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
71 UINT8 ff_alternate_horizontal_scan[64] = {
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
72 0, 1, 2, 3, 8, 9, 16, 17,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
73 10, 11, 4, 5, 6, 7, 15, 14,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
74 13, 12, 19, 18, 24, 25, 32, 33,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
75 26, 27, 20, 21, 22, 23, 28, 29,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
76 30, 31, 34, 35, 40, 41, 48, 49,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
77 42, 43, 36, 37, 38, 39, 44, 45,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
78 46, 47, 50, 51, 56, 57, 58, 59,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
79 52, 53, 54, 55, 60, 61, 62, 63,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
80 };
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
81
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
82 UINT8 ff_alternate_vertical_scan[64] = {
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
83 0, 8, 16, 24, 1, 9, 2, 10,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
84 17, 25, 32, 40, 48, 56, 57, 49,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
85 41, 33, 26, 18, 3, 11, 4, 12,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
86 19, 27, 34, 42, 50, 58, 35, 43,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
87 51, 59, 20, 28, 5, 13, 6, 14,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
88 21, 29, 36, 44, 52, 60, 37, 45,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
89 53, 61, 22, 30, 7, 15, 23, 31,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
90 38, 46, 54, 62, 39, 47, 55, 63,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
91 };
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
92
320
cda7d0857baf - ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents: 312
diff changeset
93 #ifdef SIMPLE_IDCT
cda7d0857baf - ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents: 312
diff changeset
94
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 200
diff changeset
95 /* Input permutation for the simple_idct_mmx */
190
9e0e56869d05 fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents: 174
diff changeset
96 static UINT8 simple_mmx_permutation[64]={
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 200
diff changeset
97 0x00, 0x08, 0x04, 0x09, 0x01, 0x0C, 0x05, 0x0D,
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 200
diff changeset
98 0x10, 0x18, 0x14, 0x19, 0x11, 0x1C, 0x15, 0x1D,
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 200
diff changeset
99 0x20, 0x28, 0x24, 0x29, 0x21, 0x2C, 0x25, 0x2D,
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 200
diff changeset
100 0x12, 0x1A, 0x16, 0x1B, 0x13, 0x1E, 0x17, 0x1F,
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 200
diff changeset
101 0x02, 0x0A, 0x06, 0x0B, 0x03, 0x0E, 0x07, 0x0F,
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 200
diff changeset
102 0x30, 0x38, 0x34, 0x39, 0x31, 0x3C, 0x35, 0x3D,
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 200
diff changeset
103 0x22, 0x2A, 0x26, 0x2B, 0x23, 0x2E, 0x27, 0x2F,
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 200
diff changeset
104 0x32, 0x3A, 0x36, 0x3B, 0x33, 0x3E, 0x37, 0x3F,
190
9e0e56869d05 fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents: 174
diff changeset
105 };
320
cda7d0857baf - ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents: 312
diff changeset
106 #endif
190
9e0e56869d05 fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents: 174
diff changeset
107
220
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
108 /* a*inverse[b]>>32 == a/b for all 0<=a<=65536 && 2<=b<=255 */
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
109 UINT32 inverse[256]={
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
110 0, 4294967295U,2147483648U,1431655766, 1073741824, 858993460, 715827883, 613566757,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
111 536870912, 477218589, 429496730, 390451573, 357913942, 330382100, 306783379, 286331154,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
112 268435456, 252645136, 238609295, 226050911, 214748365, 204522253, 195225787, 186737709,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
113 178956971, 171798692, 165191050, 159072863, 153391690, 148102321, 143165577, 138547333,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
114 134217728, 130150525, 126322568, 122713352, 119304648, 116080198, 113025456, 110127367,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
115 107374183, 104755300, 102261127, 99882961, 97612894, 95443718, 93368855, 91382283,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
116 89478486, 87652394, 85899346, 84215046, 82595525, 81037119, 79536432, 78090315,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
117 76695845, 75350304, 74051161, 72796056, 71582789, 70409300, 69273667, 68174085,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
118 67108864, 66076420, 65075263, 64103990, 63161284, 62245903, 61356676, 60492498,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
119 59652324, 58835169, 58040099, 57266231, 56512728, 55778797, 55063684, 54366675,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
120 53687092, 53024288, 52377650, 51746594, 51130564, 50529028, 49941481, 49367441,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
121 48806447, 48258060, 47721859, 47197443, 46684428, 46182445, 45691142, 45210183,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
122 44739243, 44278014, 43826197, 43383509, 42949673, 42524429, 42107523, 41698712,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
123 41297763, 40904451, 40518560, 40139882, 39768216, 39403370, 39045158, 38693400,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
124 38347923, 38008561, 37675152, 37347542, 37025581, 36709123, 36398028, 36092163,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
125 35791395, 35495598, 35204650, 34918434, 34636834, 34359739, 34087043, 33818641,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
126 33554432, 33294321, 33038210, 32786010, 32537632, 32292988, 32051995, 31814573,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
127 31580642, 31350127, 31122952, 30899046, 30678338, 30460761, 30246249, 30034737,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
128 29826162, 29620465, 29417585, 29217465, 29020050, 28825284, 28633116, 28443493,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
129 28256364, 28071682, 27889399, 27709467, 27531842, 27356480, 27183338, 27012373,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
130 26843546, 26676816, 26512144, 26349493, 26188825, 26030105, 25873297, 25718368,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
131 25565282, 25414008, 25264514, 25116768, 24970741, 24826401, 24683721, 24542671,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
132 24403224, 24265352, 24129030, 23994231, 23860930, 23729102, 23598722, 23469767,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
133 23342214, 23216040, 23091223, 22967740, 22845571, 22724695, 22605092, 22486740,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
134 22369622, 22253717, 22139007, 22025474, 21913099, 21801865, 21691755, 21582751,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
135 21474837, 21367997, 21262215, 21157475, 21053762, 20951060, 20849356, 20748635,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
136 20648882, 20550083, 20452226, 20355296, 20259280, 20164166, 20069941, 19976593,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
137 19884108, 19792477, 19701685, 19611723, 19522579, 19434242, 19346700, 19259944,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
138 19173962, 19088744, 19004281, 18920561, 18837576, 18755316, 18673771, 18592933,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
139 18512791, 18433337, 18354562, 18276457, 18199014, 18122225, 18046082, 17970575,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
140 17895698, 17821442, 17747799, 17674763, 17602325, 17530479, 17459217, 17388532,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
141 17318417, 17248865, 17179870, 17111424, 17043522, 16976156, 16909321, 16843010,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
142 };
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
143
200
6ab301aaa652 (commit by michael)
arpi_esp
parents: 190
diff changeset
144 /* used to skip zeros at the end */
6ab301aaa652 (commit by michael)
arpi_esp
parents: 190
diff changeset
145 UINT8 zigzag_end[64];
6ab301aaa652 (commit by michael)
arpi_esp
parents: 190
diff changeset
146
190
9e0e56869d05 fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents: 174
diff changeset
147 UINT8 permutation[64];
9e0e56869d05 fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents: 174
diff changeset
148 //UINT8 invPermutation[64];
9e0e56869d05 fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents: 174
diff changeset
149
468
1e23eae32087 Minor warning cleanup.
mellum
parents: 435
diff changeset
150 static void build_zigzag_end(void)
200
6ab301aaa652 (commit by michael)
arpi_esp
parents: 190
diff changeset
151 {
6ab301aaa652 (commit by michael)
arpi_esp
parents: 190
diff changeset
152 int lastIndex;
6ab301aaa652 (commit by michael)
arpi_esp
parents: 190
diff changeset
153 int lastIndexAfterPerm=0;
6ab301aaa652 (commit by michael)
arpi_esp
parents: 190
diff changeset
154 for(lastIndex=0; lastIndex<64; lastIndex++)
6ab301aaa652 (commit by michael)
arpi_esp
parents: 190
diff changeset
155 {
6ab301aaa652 (commit by michael)
arpi_esp
parents: 190
diff changeset
156 if(zigzag_direct[lastIndex] > lastIndexAfterPerm)
6ab301aaa652 (commit by michael)
arpi_esp
parents: 190
diff changeset
157 lastIndexAfterPerm= zigzag_direct[lastIndex];
6ab301aaa652 (commit by michael)
arpi_esp
parents: 190
diff changeset
158 zigzag_end[lastIndex]= lastIndexAfterPerm + 1;
6ab301aaa652 (commit by michael)
arpi_esp
parents: 190
diff changeset
159 }
6ab301aaa652 (commit by michael)
arpi_esp
parents: 190
diff changeset
160 }
6ab301aaa652 (commit by michael)
arpi_esp
parents: 190
diff changeset
161
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
162 void get_pixels_c(DCTELEM *block, const UINT8 *pixels, int line_size)
986e461dc072 Initial revision
glantau
parents:
diff changeset
163 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
164 DCTELEM *p;
986e461dc072 Initial revision
glantau
parents:
diff changeset
165 const UINT8 *pix;
986e461dc072 Initial revision
glantau
parents:
diff changeset
166 int i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
167
986e461dc072 Initial revision
glantau
parents:
diff changeset
168 /* read the pixels */
986e461dc072 Initial revision
glantau
parents:
diff changeset
169 p = block;
986e461dc072 Initial revision
glantau
parents:
diff changeset
170 pix = pixels;
986e461dc072 Initial revision
glantau
parents:
diff changeset
171 for(i=0;i<8;i++) {
986e461dc072 Initial revision
glantau
parents:
diff changeset
172 p[0] = pix[0];
986e461dc072 Initial revision
glantau
parents:
diff changeset
173 p[1] = pix[1];
986e461dc072 Initial revision
glantau
parents:
diff changeset
174 p[2] = pix[2];
986e461dc072 Initial revision
glantau
parents:
diff changeset
175 p[3] = pix[3];
986e461dc072 Initial revision
glantau
parents:
diff changeset
176 p[4] = pix[4];
986e461dc072 Initial revision
glantau
parents:
diff changeset
177 p[5] = pix[5];
986e461dc072 Initial revision
glantau
parents:
diff changeset
178 p[6] = pix[6];
986e461dc072 Initial revision
glantau
parents:
diff changeset
179 p[7] = pix[7];
986e461dc072 Initial revision
glantau
parents:
diff changeset
180 pix += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
181 p += 8;
986e461dc072 Initial revision
glantau
parents:
diff changeset
182 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
183 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
184
324
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
185 void diff_pixels_c(DCTELEM *block, const UINT8 *s1, const UINT8 *s2, int stride){
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
186 DCTELEM *p;
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
187 int i;
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
188
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
189 /* read the pixels */
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
190 p = block;
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
191 for(i=0;i<8;i++) {
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
192 p[0] = s1[0] - s2[0];
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
193 p[1] = s1[1] - s2[1];
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
194 p[2] = s1[2] - s2[2];
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
195 p[3] = s1[3] - s2[3];
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
196 p[4] = s1[4] - s2[4];
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
197 p[5] = s1[5] - s2[5];
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
198 p[6] = s1[6] - s2[6];
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
199 p[7] = s1[7] - s2[7];
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
200 s1 += stride;
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
201 s2 += stride;
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
202 p += 8;
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
203 }
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
204 }
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
205
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
206
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
207 void put_pixels_clamped_c(const DCTELEM *block, UINT8 *pixels, int line_size)
986e461dc072 Initial revision
glantau
parents:
diff changeset
208 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
209 const DCTELEM *p;
986e461dc072 Initial revision
glantau
parents:
diff changeset
210 UINT8 *pix;
986e461dc072 Initial revision
glantau
parents:
diff changeset
211 int i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
212 UINT8 *cm = cropTbl + MAX_NEG_CROP;
986e461dc072 Initial revision
glantau
parents:
diff changeset
213
986e461dc072 Initial revision
glantau
parents:
diff changeset
214 /* read the pixels */
986e461dc072 Initial revision
glantau
parents:
diff changeset
215 p = block;
986e461dc072 Initial revision
glantau
parents:
diff changeset
216 pix = pixels;
986e461dc072 Initial revision
glantau
parents:
diff changeset
217 for(i=0;i<8;i++) {
986e461dc072 Initial revision
glantau
parents:
diff changeset
218 pix[0] = cm[p[0]];
986e461dc072 Initial revision
glantau
parents:
diff changeset
219 pix[1] = cm[p[1]];
986e461dc072 Initial revision
glantau
parents:
diff changeset
220 pix[2] = cm[p[2]];
986e461dc072 Initial revision
glantau
parents:
diff changeset
221 pix[3] = cm[p[3]];
986e461dc072 Initial revision
glantau
parents:
diff changeset
222 pix[4] = cm[p[4]];
986e461dc072 Initial revision
glantau
parents:
diff changeset
223 pix[5] = cm[p[5]];
986e461dc072 Initial revision
glantau
parents:
diff changeset
224 pix[6] = cm[p[6]];
986e461dc072 Initial revision
glantau
parents:
diff changeset
225 pix[7] = cm[p[7]];
986e461dc072 Initial revision
glantau
parents:
diff changeset
226 pix += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
227 p += 8;
986e461dc072 Initial revision
glantau
parents:
diff changeset
228 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
229 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
230
986e461dc072 Initial revision
glantau
parents:
diff changeset
231 void add_pixels_clamped_c(const DCTELEM *block, UINT8 *pixels, int line_size)
986e461dc072 Initial revision
glantau
parents:
diff changeset
232 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
233 const DCTELEM *p;
986e461dc072 Initial revision
glantau
parents:
diff changeset
234 UINT8 *pix;
986e461dc072 Initial revision
glantau
parents:
diff changeset
235 int i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
236 UINT8 *cm = cropTbl + MAX_NEG_CROP;
986e461dc072 Initial revision
glantau
parents:
diff changeset
237
986e461dc072 Initial revision
glantau
parents:
diff changeset
238 /* read the pixels */
986e461dc072 Initial revision
glantau
parents:
diff changeset
239 p = block;
986e461dc072 Initial revision
glantau
parents:
diff changeset
240 pix = pixels;
986e461dc072 Initial revision
glantau
parents:
diff changeset
241 for(i=0;i<8;i++) {
986e461dc072 Initial revision
glantau
parents:
diff changeset
242 pix[0] = cm[pix[0] + p[0]];
986e461dc072 Initial revision
glantau
parents:
diff changeset
243 pix[1] = cm[pix[1] + p[1]];
986e461dc072 Initial revision
glantau
parents:
diff changeset
244 pix[2] = cm[pix[2] + p[2]];
986e461dc072 Initial revision
glantau
parents:
diff changeset
245 pix[3] = cm[pix[3] + p[3]];
986e461dc072 Initial revision
glantau
parents:
diff changeset
246 pix[4] = cm[pix[4] + p[4]];
986e461dc072 Initial revision
glantau
parents:
diff changeset
247 pix[5] = cm[pix[5] + p[5]];
986e461dc072 Initial revision
glantau
parents:
diff changeset
248 pix[6] = cm[pix[6] + p[6]];
986e461dc072 Initial revision
glantau
parents:
diff changeset
249 pix[7] = cm[pix[7] + p[7]];
986e461dc072 Initial revision
glantau
parents:
diff changeset
250 pix += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
251 p += 8;
986e461dc072 Initial revision
glantau
parents:
diff changeset
252 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
253 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
254
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
255 #if 0
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
256
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
257 #define PIXOP2(OPNAME, OP) \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
258 void OPNAME ## _pixels(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
259 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
260 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
261 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
262 OP(*((uint64_t*)block), LD64(pixels));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
263 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
264 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
265 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
266 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
267 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
268 void OPNAME ## _no_rnd_pixels_x2(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
269 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
270 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
271 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
272 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
273 const uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
274 OP(*((uint64_t*)block), (a&b) + (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
275 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
276 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
277 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
278 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
279 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
280 void OPNAME ## _pixels_x2(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
281 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
282 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
283 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
284 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
285 const uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
286 OP(*((uint64_t*)block), (a|b) - (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
287 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
288 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
289 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
290 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
291 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
292 void OPNAME ## _no_rnd_pixels_y2(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
293 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
294 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
295 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
296 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
297 const uint64_t b= LD64(pixels+line_size);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
298 OP(*((uint64_t*)block), (a&b) + (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
299 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
300 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
301 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
302 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
303 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
304 void OPNAME ## _pixels_y2(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
305 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
306 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
307 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
308 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
309 const uint64_t b= LD64(pixels+line_size);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
310 OP(*((uint64_t*)block), (a|b) - (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
311 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
312 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
313 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
314 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
315 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
316 void OPNAME ## _pixels_xy2(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
317 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
318 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
319 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
320 const uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
321 uint64_t l0= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
322 + (b&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
323 + 0x0202020202020202ULL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
324 uint64_t h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
325 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
326 uint64_t l1,h1;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
327 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
328 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
329 for(i=0; i<h; i+=2){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
330 uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
331 uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
332 l1= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
333 + (b&0x0303030303030303ULL);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
334 h1= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
335 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
336 OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
337 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
338 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
339 a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
340 b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
341 l0= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
342 + (b&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
343 + 0x0202020202020202ULL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
344 h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
345 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
346 OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
347 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
348 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
349 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
350 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
351 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
352 void OPNAME ## _no_rnd_pixels_xy2(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
353 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
354 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
355 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
356 const uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
357 uint64_t l0= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
358 + (b&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
359 + 0x0101010101010101ULL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
360 uint64_t h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
361 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
362 uint64_t l1,h1;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
363 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
364 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
365 for(i=0; i<h; i+=2){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
366 uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
367 uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
368 l1= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
369 + (b&0x0303030303030303ULL);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
370 h1= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
371 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
372 OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
373 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
374 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
375 a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
376 b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
377 l0= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
378 + (b&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
379 + 0x0101010101010101ULL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
380 h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
381 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
382 OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
383 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
384 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
385 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
386 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
387 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
388 void (*OPNAME ## _pixels_tab[4])(uint8_t *block, const uint8_t *pixels, int line_size, int h) = {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
389 OPNAME ## _pixels,\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
390 OPNAME ## _pixels_x2,\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
391 OPNAME ## _pixels_y2,\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
392 OPNAME ## _pixels_xy2,\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
393 };\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
394 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
395 void (*OPNAME ## _no_rnd_pixels_tab[4])(uint8_t *block, const uint8_t *pixels, int line_size, int h) = {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
396 OPNAME ## _pixels,\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
397 OPNAME ## _no_rnd_pixels_x2,\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
398 OPNAME ## _no_rnd_pixels_y2,\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
399 OPNAME ## _no_rnd_pixels_xy2,\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
400 };
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
401
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
402 #define op_avg(a, b) a = ( ((a)|(b)) - ((((a)^(b))&0xFEFEFEFEFEFEFEFEULL)>>1) )
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
403 #else // 64 bit variant
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
404
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
405 #define PIXOP2(OPNAME, OP) \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
406 void OPNAME ## _pixels(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
407 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
408 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
409 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
410 OP(*((uint32_t*)(block )), LD32(pixels ));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
411 OP(*((uint32_t*)(block+4)), LD32(pixels+4));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
412 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
413 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
414 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
415 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
416 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
417 void OPNAME ## _no_rnd_pixels_x2(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
418 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
419 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
420 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
421 int j;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
422 for(j=0; j<2; j++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
423 const uint32_t a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
424 const uint32_t b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
425 OP(*((uint32_t*)block), (a&b) + (((a^b)&0xFEFEFEFEUL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
426 pixels+=4;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
427 block +=4;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
428 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
429 pixels+=line_size-8;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
430 block +=line_size-8;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
431 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
432 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
433 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
434 void OPNAME ## _pixels_x2(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
435 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
436 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
437 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
438 int j;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
439 for(j=0; j<2; j++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
440 const uint32_t a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
441 const uint32_t b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
442 OP(*((uint32_t*)block), (a|b) - (((a^b)&0xFEFEFEFEUL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
443 pixels+=4;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
444 block +=4;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
445 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
446 pixels+=line_size-8;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
447 block +=line_size-8;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
448 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
449 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
450 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
451 void OPNAME ## _no_rnd_pixels_y2(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
452 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
453 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
454 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
455 int j;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
456 for(j=0; j<2; j++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
457 const uint32_t a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
458 const uint32_t b= LD32(pixels+line_size);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
459 OP(*((uint32_t*)block), (a&b) + (((a^b)&0xFEFEFEFEUL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
460 pixels+=4;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
461 block +=4;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
462 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
463 pixels+=line_size-8;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
464 block +=line_size-8;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
465 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
466 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
467 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
468 void OPNAME ## _pixels_y2(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
469 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
470 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
471 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
472 int j;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
473 for(j=0; j<2; j++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
474 const uint32_t a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
475 const uint32_t b= LD32(pixels+line_size);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
476 OP(*((uint32_t*)block), (a|b) - (((a^b)&0xFEFEFEFEUL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
477 pixels+=4;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
478 block +=4;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
479 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
480 pixels+=line_size-8;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
481 block +=line_size-8;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
482 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
483 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
484 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
485 void OPNAME ## _pixels_xy2(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
486 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
487 int j;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
488 for(j=0; j<2; j++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
489 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
490 const uint32_t a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
491 const uint32_t b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
492 uint32_t l0= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
493 + (b&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
494 + 0x02020202UL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
495 uint32_t h0= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
496 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
497 uint32_t l1,h1;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
498 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
499 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
500 for(i=0; i<h; i+=2){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
501 uint32_t a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
502 uint32_t b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
503 l1= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
504 + (b&0x03030303UL);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
505 h1= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
506 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
507 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
508 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
509 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
510 a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
511 b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
512 l0= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
513 + (b&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
514 + 0x02020202UL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
515 h0= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
516 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
517 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
518 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
519 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
520 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
521 pixels+=4-line_size*(h+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
522 block +=4-line_size*h;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
523 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
524 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
525 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
526 void OPNAME ## _no_rnd_pixels_xy2(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
527 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
528 int j;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
529 for(j=0; j<2; j++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
530 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
531 const uint32_t a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
532 const uint32_t b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
533 uint32_t l0= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
534 + (b&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
535 + 0x01010101UL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
536 uint32_t h0= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
537 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
538 uint32_t l1,h1;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
539 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
540 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
541 for(i=0; i<h; i+=2){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
542 uint32_t a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
543 uint32_t b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
544 l1= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
545 + (b&0x03030303UL);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
546 h1= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
547 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
548 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
549 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
550 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
551 a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
552 b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
553 l0= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
554 + (b&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
555 + 0x01010101UL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
556 h0= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
557 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
558 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
559 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
560 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
561 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
562 pixels+=4-line_size*(h+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
563 block +=4-line_size*h;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
564 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
565 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
566 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
567 void (*OPNAME ## _pixels_tab[4])(uint8_t *block, const uint8_t *pixels, int line_size, int h) = {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
568 OPNAME ## _pixels,\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
569 OPNAME ## _pixels_x2,\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
570 OPNAME ## _pixels_y2,\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
571 OPNAME ## _pixels_xy2,\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
572 };\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
573 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
574 void (*OPNAME ## _no_rnd_pixels_tab[4])(uint8_t *block, const uint8_t *pixels, int line_size, int h) = {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
575 OPNAME ## _pixels,\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
576 OPNAME ## _no_rnd_pixels_x2,\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
577 OPNAME ## _no_rnd_pixels_y2,\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
578 OPNAME ## _no_rnd_pixels_xy2,\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
579 };
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
580 #define op_avg(a, b) a = ( ((a)|(b)) - ((((a)^(b))&0xFEFEFEFEUL)>>1) )
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
581 #endif
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
582
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
583 #define op_put(a, b) a = b
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
584
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
585 PIXOP2(avg, op_avg)
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
586 PIXOP2(put, op_put)
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
587 #undef op_avg
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
588 #undef op_put
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
589
403
2c3e25f4c496 removed unused stuff - added dsputil_set_bit_exact() support for easier testing
glantau
parents: 398
diff changeset
590 #if 0
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
591 /* FIXME this stuff could be removed as its ot really used anymore */
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
592 #define PIXOP(BTYPE, OPNAME, OP, INCR) \
986e461dc072 Initial revision
glantau
parents:
diff changeset
593 \
986e461dc072 Initial revision
glantau
parents:
diff changeset
594 static void OPNAME ## _pixels(BTYPE *block, const UINT8 *pixels, int line_size, int h) \
986e461dc072 Initial revision
glantau
parents:
diff changeset
595 { \
986e461dc072 Initial revision
glantau
parents:
diff changeset
596 BTYPE *p; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
597 const UINT8 *pix; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
598 \
986e461dc072 Initial revision
glantau
parents:
diff changeset
599 p = block; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
600 pix = pixels; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
601 do { \
986e461dc072 Initial revision
glantau
parents:
diff changeset
602 OP(p[0], pix[0]); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
603 OP(p[1], pix[1]); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
604 OP(p[2], pix[2]); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
605 OP(p[3], pix[3]); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
606 OP(p[4], pix[4]); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
607 OP(p[5], pix[5]); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
608 OP(p[6], pix[6]); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
609 OP(p[7], pix[7]); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
610 pix += line_size; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
611 p += INCR; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
612 } while (--h);; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
613 } \
986e461dc072 Initial revision
glantau
parents:
diff changeset
614 \
986e461dc072 Initial revision
glantau
parents:
diff changeset
615 static void OPNAME ## _pixels_x2(BTYPE *block, const UINT8 *pixels, int line_size, int h) \
986e461dc072 Initial revision
glantau
parents:
diff changeset
616 { \
986e461dc072 Initial revision
glantau
parents:
diff changeset
617 BTYPE *p; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
618 const UINT8 *pix; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
619 \
986e461dc072 Initial revision
glantau
parents:
diff changeset
620 p = block; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
621 pix = pixels; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
622 do { \
986e461dc072 Initial revision
glantau
parents:
diff changeset
623 OP(p[0], avg2(pix[0], pix[1])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
624 OP(p[1], avg2(pix[1], pix[2])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
625 OP(p[2], avg2(pix[2], pix[3])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
626 OP(p[3], avg2(pix[3], pix[4])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
627 OP(p[4], avg2(pix[4], pix[5])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
628 OP(p[5], avg2(pix[5], pix[6])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
629 OP(p[6], avg2(pix[6], pix[7])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
630 OP(p[7], avg2(pix[7], pix[8])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
631 pix += line_size; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
632 p += INCR; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
633 } while (--h); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
634 } \
986e461dc072 Initial revision
glantau
parents:
diff changeset
635 \
986e461dc072 Initial revision
glantau
parents:
diff changeset
636 static void OPNAME ## _pixels_y2(BTYPE *block, const UINT8 *pixels, int line_size, int h) \
986e461dc072 Initial revision
glantau
parents:
diff changeset
637 { \
986e461dc072 Initial revision
glantau
parents:
diff changeset
638 BTYPE *p; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
639 const UINT8 *pix; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
640 const UINT8 *pix1; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
641 \
986e461dc072 Initial revision
glantau
parents:
diff changeset
642 p = block; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
643 pix = pixels; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
644 pix1 = pixels + line_size; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
645 do { \
986e461dc072 Initial revision
glantau
parents:
diff changeset
646 OP(p[0], avg2(pix[0], pix1[0])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
647 OP(p[1], avg2(pix[1], pix1[1])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
648 OP(p[2], avg2(pix[2], pix1[2])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
649 OP(p[3], avg2(pix[3], pix1[3])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
650 OP(p[4], avg2(pix[4], pix1[4])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
651 OP(p[5], avg2(pix[5], pix1[5])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
652 OP(p[6], avg2(pix[6], pix1[6])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
653 OP(p[7], avg2(pix[7], pix1[7])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
654 pix += line_size; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
655 pix1 += line_size; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
656 p += INCR; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
657 } while(--h); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
658 } \
986e461dc072 Initial revision
glantau
parents:
diff changeset
659 \
986e461dc072 Initial revision
glantau
parents:
diff changeset
660 static void OPNAME ## _pixels_xy2(BTYPE *block, const UINT8 *pixels, int line_size, int h) \
986e461dc072 Initial revision
glantau
parents:
diff changeset
661 { \
986e461dc072 Initial revision
glantau
parents:
diff changeset
662 BTYPE *p; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
663 const UINT8 *pix; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
664 const UINT8 *pix1; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
665 \
986e461dc072 Initial revision
glantau
parents:
diff changeset
666 p = block; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
667 pix = pixels; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
668 pix1 = pixels + line_size; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
669 do { \
986e461dc072 Initial revision
glantau
parents:
diff changeset
670 OP(p[0], avg4(pix[0], pix[1], pix1[0], pix1[1])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
671 OP(p[1], avg4(pix[1], pix[2], pix1[1], pix1[2])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
672 OP(p[2], avg4(pix[2], pix[3], pix1[2], pix1[3])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
673 OP(p[3], avg4(pix[3], pix[4], pix1[3], pix1[4])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
674 OP(p[4], avg4(pix[4], pix[5], pix1[4], pix1[5])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
675 OP(p[5], avg4(pix[5], pix[6], pix1[5], pix1[6])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
676 OP(p[6], avg4(pix[6], pix[7], pix1[6], pix1[7])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
677 OP(p[7], avg4(pix[7], pix[8], pix1[7], pix1[8])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
678 pix += line_size; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
679 pix1 += line_size; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
680 p += INCR; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
681 } while(--h); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
682 } \
986e461dc072 Initial revision
glantau
parents:
diff changeset
683 \
986e461dc072 Initial revision
glantau
parents:
diff changeset
684 void (*OPNAME ## _pixels_tab[4])(BTYPE *block, const UINT8 *pixels, int line_size, int h) = { \
986e461dc072 Initial revision
glantau
parents:
diff changeset
685 OPNAME ## _pixels, \
986e461dc072 Initial revision
glantau
parents:
diff changeset
686 OPNAME ## _pixels_x2, \
986e461dc072 Initial revision
glantau
parents:
diff changeset
687 OPNAME ## _pixels_y2, \
986e461dc072 Initial revision
glantau
parents:
diff changeset
688 OPNAME ## _pixels_xy2, \
986e461dc072 Initial revision
glantau
parents:
diff changeset
689 };
986e461dc072 Initial revision
glantau
parents:
diff changeset
690
986e461dc072 Initial revision
glantau
parents:
diff changeset
691 /* rounding primitives */
986e461dc072 Initial revision
glantau
parents:
diff changeset
692 #define avg2(a,b) ((a+b+1)>>1)
986e461dc072 Initial revision
glantau
parents:
diff changeset
693 #define avg4(a,b,c,d) ((a+b+c+d+2)>>2)
986e461dc072 Initial revision
glantau
parents:
diff changeset
694
986e461dc072 Initial revision
glantau
parents:
diff changeset
695 #define op_avg(a, b) a = avg2(a, b)
986e461dc072 Initial revision
glantau
parents:
diff changeset
696 #define op_sub(a, b) a -= b
986e461dc072 Initial revision
glantau
parents:
diff changeset
697
986e461dc072 Initial revision
glantau
parents:
diff changeset
698 PIXOP(DCTELEM, sub, op_sub, 8)
986e461dc072 Initial revision
glantau
parents:
diff changeset
699
986e461dc072 Initial revision
glantau
parents:
diff changeset
700 /* not rounding primitives */
986e461dc072 Initial revision
glantau
parents:
diff changeset
701 #undef avg2
986e461dc072 Initial revision
glantau
parents:
diff changeset
702 #undef avg4
986e461dc072 Initial revision
glantau
parents:
diff changeset
703 #define avg2(a,b) ((a+b)>>1)
986e461dc072 Initial revision
glantau
parents:
diff changeset
704 #define avg4(a,b,c,d) ((a+b+c+d+1)>>2)
986e461dc072 Initial revision
glantau
parents:
diff changeset
705
986e461dc072 Initial revision
glantau
parents:
diff changeset
706 /* motion estimation */
986e461dc072 Initial revision
glantau
parents:
diff changeset
707
986e461dc072 Initial revision
glantau
parents:
diff changeset
708 #undef avg2
986e461dc072 Initial revision
glantau
parents:
diff changeset
709 #undef avg4
403
2c3e25f4c496 removed unused stuff - added dsputil_set_bit_exact() support for easier testing
glantau
parents: 398
diff changeset
710 #endif
2c3e25f4c496 removed unused stuff - added dsputil_set_bit_exact() support for easier testing
glantau
parents: 398
diff changeset
711
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
712 #define avg2(a,b) ((a+b+1)>>1)
986e461dc072 Initial revision
glantau
parents:
diff changeset
713 #define avg4(a,b,c,d) ((a+b+c+d+2)>>2)
986e461dc072 Initial revision
glantau
parents:
diff changeset
714
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
715 static void gmc1_c(UINT8 *dst, UINT8 *src, int srcStride, int h, int x16, int y16, int rounder)
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
716 {
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
717 const int A=(16-x16)*(16-y16);
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
718 const int B=( x16)*(16-y16);
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
719 const int C=(16-x16)*( y16);
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
720 const int D=( x16)*( y16);
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
721 int i;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
722 rounder= 128 - rounder;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
723
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
724 for(i=0; i<h; i++)
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
725 {
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
726 dst[0]= (A*src[0] + B*src[1] + C*src[srcStride+0] + D*src[srcStride+1] + rounder)>>8;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
727 dst[1]= (A*src[1] + B*src[2] + C*src[srcStride+1] + D*src[srcStride+2] + rounder)>>8;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
728 dst[2]= (A*src[2] + B*src[3] + C*src[srcStride+2] + D*src[srcStride+3] + rounder)>>8;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
729 dst[3]= (A*src[3] + B*src[4] + C*src[srcStride+3] + D*src[srcStride+4] + rounder)>>8;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
730 dst[4]= (A*src[4] + B*src[5] + C*src[srcStride+4] + D*src[srcStride+5] + rounder)>>8;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
731 dst[5]= (A*src[5] + B*src[6] + C*src[srcStride+5] + D*src[srcStride+6] + rounder)>>8;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
732 dst[6]= (A*src[6] + B*src[7] + C*src[srcStride+6] + D*src[srcStride+7] + rounder)>>8;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
733 dst[7]= (A*src[7] + B*src[8] + C*src[srcStride+7] + D*src[srcStride+8] + rounder)>>8;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
734 dst+= srcStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
735 src+= srcStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
736 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
737 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
738
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
739 static void qpel_h_lowpass(UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int h, int r)
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
740 {
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
741 UINT8 *cm = cropTbl + MAX_NEG_CROP;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
742 int i;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
743 for(i=0; i<h; i++)
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
744 {
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
745 dst[0]= cm[(((src[0]+src[1])*20 - (src[0]+src[2])*6 + (src[1]+src[3])*3 - (src[2]+src[4]) + r)>>5)];
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
746 dst[1]= cm[(((src[1]+src[2])*20 - (src[0]+src[3])*6 + (src[0]+src[4])*3 - (src[1]+src[5]) + r)>>5)];
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
747 dst[2]= cm[(((src[2]+src[3])*20 - (src[1]+src[4])*6 + (src[0]+src[5])*3 - (src[0]+src[6]) + r)>>5)];
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
748 dst[3]= cm[(((src[3]+src[4])*20 - (src[2]+src[5])*6 + (src[1]+src[6])*3 - (src[0]+src[7]) + r)>>5)];
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
749 dst[4]= cm[(((src[4]+src[5])*20 - (src[3]+src[6])*6 + (src[2]+src[7])*3 - (src[1]+src[8]) + r)>>5)];
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
750 dst[5]= cm[(((src[5]+src[6])*20 - (src[4]+src[7])*6 + (src[3]+src[8])*3 - (src[2]+src[8]) + r)>>5)];
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
751 dst[6]= cm[(((src[6]+src[7])*20 - (src[5]+src[8])*6 + (src[4]+src[8])*3 - (src[3]+src[7]) + r)>>5)];
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
752 dst[7]= cm[(((src[7]+src[8])*20 - (src[6]+src[8])*6 + (src[5]+src[7])*3 - (src[4]+src[6]) + r)>>5)];
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
753 dst+=dstStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
754 src+=srcStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
755 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
756 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
757
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
758 static void qpel_v_lowpass(UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int w, int r)
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
759 {
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
760 UINT8 *cm = cropTbl + MAX_NEG_CROP;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
761 int i;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
762 for(i=0; i<w; i++)
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
763 {
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
764 const int src0= src[0*srcStride];
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
765 const int src1= src[1*srcStride];
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
766 const int src2= src[2*srcStride];
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
767 const int src3= src[3*srcStride];
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
768 const int src4= src[4*srcStride];
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
769 const int src5= src[5*srcStride];
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
770 const int src6= src[6*srcStride];
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
771 const int src7= src[7*srcStride];
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
772 const int src8= src[8*srcStride];
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
773 dst[0*dstStride]= cm[(((src0+src1)*20 - (src0+src2)*6 + (src1+src3)*3 - (src2+src4) + r)>>5)];
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
774 dst[1*dstStride]= cm[(((src1+src2)*20 - (src0+src3)*6 + (src0+src4)*3 - (src1+src5) + r)>>5)];
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
775 dst[2*dstStride]= cm[(((src2+src3)*20 - (src1+src4)*6 + (src0+src5)*3 - (src0+src6) + r)>>5)];
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
776 dst[3*dstStride]= cm[(((src3+src4)*20 - (src2+src5)*6 + (src1+src6)*3 - (src0+src7) + r)>>5)];
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
777 dst[4*dstStride]= cm[(((src4+src5)*20 - (src3+src6)*6 + (src2+src7)*3 - (src1+src8) + r)>>5)];
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
778 dst[5*dstStride]= cm[(((src5+src6)*20 - (src4+src7)*6 + (src3+src8)*3 - (src2+src8) + r)>>5)];
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
779 dst[6*dstStride]= cm[(((src6+src7)*20 - (src5+src8)*6 + (src4+src8)*3 - (src3+src7) + r)>>5)];
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
780 dst[7*dstStride]= cm[(((src7+src8)*20 - (src6+src8)*6 + (src5+src7)*3 - (src4+src6) + r)>>5)];
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
781 dst++;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
782 src++;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
783 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
784 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
785
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
786 static inline void put_block(UINT8 *dst, UINT8 *src, int dstStride, int srcStride)
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
787 {
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
788 int i;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
789 for(i=0; i<8; i++)
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
790 {
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
791 dst[0]= src[0];
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
792 dst[1]= src[1];
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
793 dst[2]= src[2];
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
794 dst[3]= src[3];
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
795 dst[4]= src[4];
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
796 dst[5]= src[5];
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
797 dst[6]= src[6];
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
798 dst[7]= src[7];
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
799 dst+=dstStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
800 src+=srcStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
801 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
802 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
803
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
804 static inline void avg2_block(UINT8 *dst, UINT8 *src1, UINT8 *src2, int dstStride, int srcStride, int r)
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
805 {
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
806 int i;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
807 for(i=0; i<8; i++)
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
808 {
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
809 dst[0]= (src1[0] + src2[0] + r)>>1;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
810 dst[1]= (src1[1] + src2[1] + r)>>1;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
811 dst[2]= (src1[2] + src2[2] + r)>>1;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
812 dst[3]= (src1[3] + src2[3] + r)>>1;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
813 dst[4]= (src1[4] + src2[4] + r)>>1;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
814 dst[5]= (src1[5] + src2[5] + r)>>1;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
815 dst[6]= (src1[6] + src2[6] + r)>>1;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
816 dst[7]= (src1[7] + src2[7] + r)>>1;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
817 dst+=dstStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
818 src1+=srcStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
819 src2+=8;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
820 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
821 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
822
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
823 static inline void avg4_block(UINT8 *dst, UINT8 *src1, UINT8 *src2, UINT8 *src3, UINT8 *src4, int dstStride, int srcStride, int r)
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
824 {
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
825 int i;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
826 for(i=0; i<8; i++)
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
827 {
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
828 dst[0]= (src1[0] + src2[0] + src3[0] + src4[0] + r)>>2;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
829 dst[1]= (src1[1] + src2[1] + src3[1] + src4[1] + r)>>2;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
830 dst[2]= (src1[2] + src2[2] + src3[2] + src4[2] + r)>>2;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
831 dst[3]= (src1[3] + src2[3] + src3[3] + src4[3] + r)>>2;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
832 dst[4]= (src1[4] + src2[4] + src3[4] + src4[4] + r)>>2;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
833 dst[5]= (src1[5] + src2[5] + src3[5] + src4[5] + r)>>2;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
834 dst[6]= (src1[6] + src2[6] + src3[6] + src4[6] + r)>>2;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
835 dst[7]= (src1[7] + src2[7] + src3[7] + src4[7] + r)>>2;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
836 dst+=dstStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
837 src1+=srcStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
838 src2+=8;
256
4c1cec7c3c7c q-pel mc fixed
michaelni
parents: 255
diff changeset
839 src3+=8;
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
840 src4+=8;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
841 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
842 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
843
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
844 #define QPEL_MC(r, name) \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
845 static void qpel_mc00_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
846 {\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
847 put_block(dst, src, dstStride, srcStride);\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
848 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
849 \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
850 static void qpel_mc10_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
851 {\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
852 UINT8 half[64];\
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
853 qpel_h_lowpass(half, src, 8, srcStride, 8, 16-r);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
854 avg2_block(dst, src, half, dstStride, srcStride, 1-r);\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
855 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
856 \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
857 static void qpel_mc20_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
858 {\
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
859 qpel_h_lowpass(dst, src, dstStride, srcStride, 8, 16-r);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
860 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
861 \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
862 static void qpel_mc30_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
863 {\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
864 UINT8 half[64];\
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
865 qpel_h_lowpass(half, src, 8, srcStride, 8, 16-r);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
866 avg2_block(dst, src+1, half, dstStride, srcStride, 1-r);\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
867 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
868 \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
869 static void qpel_mc01_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
870 {\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
871 UINT8 half[64];\
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
872 qpel_v_lowpass(half, src, 8, srcStride, 8, 16-r);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
873 avg2_block(dst, src, half, dstStride, srcStride, 1-r);\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
874 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
875 \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
876 static void qpel_mc02_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
877 {\
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
878 qpel_v_lowpass(dst, src, dstStride, srcStride, 8, 16-r);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
879 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
880 \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
881 static void qpel_mc03_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
882 {\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
883 UINT8 half[64];\
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
884 qpel_v_lowpass(half, src, 8, srcStride, 8, 16-r);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
885 avg2_block(dst, src+srcStride, half, dstStride, srcStride, 1-r);\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
886 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
887 static void qpel_mc11_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
888 {\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
889 UINT8 halfH[72];\
256
4c1cec7c3c7c q-pel mc fixed
michaelni
parents: 255
diff changeset
890 UINT8 halfV[64];\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
891 UINT8 halfHV[64];\
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
892 qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
893 qpel_v_lowpass(halfV, src, 8, srcStride, 8, 16-r);\
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
894 qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
895 avg4_block(dst, src, halfH, halfV, halfHV, dstStride, srcStride, 2-r);\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
896 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
897 static void qpel_mc31_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
898 {\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
899 UINT8 halfH[72];\
256
4c1cec7c3c7c q-pel mc fixed
michaelni
parents: 255
diff changeset
900 UINT8 halfV[64];\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
901 UINT8 halfHV[64];\
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
902 qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
903 qpel_v_lowpass(halfV, src+1, 8, srcStride, 8, 16-r);\
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
904 qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
905 avg4_block(dst, src+1, halfH, halfV, halfHV, dstStride, srcStride, 2-r);\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
906 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
907 static void qpel_mc13_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
908 {\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
909 UINT8 halfH[72];\
256
4c1cec7c3c7c q-pel mc fixed
michaelni
parents: 255
diff changeset
910 UINT8 halfV[64];\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
911 UINT8 halfHV[64];\
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
912 qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
913 qpel_v_lowpass(halfV, src, 8, srcStride, 8, 16-r);\
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
914 qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\
256
4c1cec7c3c7c q-pel mc fixed
michaelni
parents: 255
diff changeset
915 avg4_block(dst, src+srcStride, halfH+8, halfV, halfHV, dstStride, srcStride, 2-r);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
916 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
917 static void qpel_mc33_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
918 {\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
919 UINT8 halfH[72];\
256
4c1cec7c3c7c q-pel mc fixed
michaelni
parents: 255
diff changeset
920 UINT8 halfV[64];\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
921 UINT8 halfHV[64];\
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
922 qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
923 qpel_v_lowpass(halfV, src+1, 8, srcStride, 8, 16-r);\
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
924 qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\
256
4c1cec7c3c7c q-pel mc fixed
michaelni
parents: 255
diff changeset
925 avg4_block(dst, src+srcStride+1, halfH+8, halfV, halfHV, dstStride, srcStride, 2-r);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
926 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
927 static void qpel_mc21_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
928 {\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
929 UINT8 halfH[72];\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
930 UINT8 halfHV[64];\
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
931 qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
932 qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
933 avg2_block(dst, halfH, halfHV, dstStride, 8, 1-r);\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
934 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
935 static void qpel_mc23_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
936 {\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
937 UINT8 halfH[72];\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
938 UINT8 halfHV[64];\
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
939 qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
940 qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
941 avg2_block(dst, halfH+8, halfHV, dstStride, 8, 1-r);\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
942 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
943 static void qpel_mc12_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
944 {\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
945 UINT8 halfH[72];\
256
4c1cec7c3c7c q-pel mc fixed
michaelni
parents: 255
diff changeset
946 UINT8 halfV[64];\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
947 UINT8 halfHV[64];\
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
948 qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
949 qpel_v_lowpass(halfV, src, 8, srcStride, 8, 16-r);\
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
950 qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\
256
4c1cec7c3c7c q-pel mc fixed
michaelni
parents: 255
diff changeset
951 avg2_block(dst, halfV, halfHV, dstStride, 8, 1-r);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
952 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
953 static void qpel_mc32_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
954 {\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
955 UINT8 halfH[72];\
256
4c1cec7c3c7c q-pel mc fixed
michaelni
parents: 255
diff changeset
956 UINT8 halfV[64];\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
957 UINT8 halfHV[64];\
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
958 qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
959 qpel_v_lowpass(halfV, src+1, 8, srcStride, 8, 16-r);\
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
960 qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\
256
4c1cec7c3c7c q-pel mc fixed
michaelni
parents: 255
diff changeset
961 avg2_block(dst, halfV, halfHV, dstStride, 8, 1-r);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
962 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
963 static void qpel_mc22_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
964 {\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
965 UINT8 halfH[72];\
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
966 qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
967 qpel_v_lowpass(dst, halfH, dstStride, 8, 8, 16-r);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
968 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
969 qpel_mc_func qpel_mc ## name ## _tab[16]={ \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
970 qpel_mc00_c ## name, \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
971 qpel_mc10_c ## name, \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
972 qpel_mc20_c ## name, \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
973 qpel_mc30_c ## name, \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
974 qpel_mc01_c ## name, \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
975 qpel_mc11_c ## name, \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
976 qpel_mc21_c ## name, \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
977 qpel_mc31_c ## name, \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
978 qpel_mc02_c ## name, \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
979 qpel_mc12_c ## name, \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
980 qpel_mc22_c ## name, \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
981 qpel_mc32_c ## name, \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
982 qpel_mc03_c ## name, \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
983 qpel_mc13_c ## name, \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
984 qpel_mc23_c ## name, \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
985 qpel_mc33_c ## name, \
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
986 };
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
987
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
988 QPEL_MC(0, _rnd)
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
989 QPEL_MC(1, _no_rnd)
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
990
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
991 int pix_abs16x16_c(UINT8 *pix1, UINT8 *pix2, int line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
992 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
993 int s, i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
994
986e461dc072 Initial revision
glantau
parents:
diff changeset
995 s = 0;
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
996 for(i=0;i<16;i++) {
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
997 s += abs(pix1[0] - pix2[0]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
998 s += abs(pix1[1] - pix2[1]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
999 s += abs(pix1[2] - pix2[2]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1000 s += abs(pix1[3] - pix2[3]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1001 s += abs(pix1[4] - pix2[4]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1002 s += abs(pix1[5] - pix2[5]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1003 s += abs(pix1[6] - pix2[6]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1004 s += abs(pix1[7] - pix2[7]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1005 s += abs(pix1[8] - pix2[8]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1006 s += abs(pix1[9] - pix2[9]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1007 s += abs(pix1[10] - pix2[10]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1008 s += abs(pix1[11] - pix2[11]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1009 s += abs(pix1[12] - pix2[12]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1010 s += abs(pix1[13] - pix2[13]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1011 s += abs(pix1[14] - pix2[14]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1012 s += abs(pix1[15] - pix2[15]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1013 pix1 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1014 pix2 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1015 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
1016 return s;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1017 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
1018
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1019 int pix_abs16x16_x2_c(UINT8 *pix1, UINT8 *pix2, int line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1020 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
1021 int s, i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1022
986e461dc072 Initial revision
glantau
parents:
diff changeset
1023 s = 0;
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1024 for(i=0;i<16;i++) {
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1025 s += abs(pix1[0] - avg2(pix2[0], pix2[1]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1026 s += abs(pix1[1] - avg2(pix2[1], pix2[2]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1027 s += abs(pix1[2] - avg2(pix2[2], pix2[3]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1028 s += abs(pix1[3] - avg2(pix2[3], pix2[4]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1029 s += abs(pix1[4] - avg2(pix2[4], pix2[5]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1030 s += abs(pix1[5] - avg2(pix2[5], pix2[6]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1031 s += abs(pix1[6] - avg2(pix2[6], pix2[7]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1032 s += abs(pix1[7] - avg2(pix2[7], pix2[8]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1033 s += abs(pix1[8] - avg2(pix2[8], pix2[9]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1034 s += abs(pix1[9] - avg2(pix2[9], pix2[10]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1035 s += abs(pix1[10] - avg2(pix2[10], pix2[11]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1036 s += abs(pix1[11] - avg2(pix2[11], pix2[12]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1037 s += abs(pix1[12] - avg2(pix2[12], pix2[13]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1038 s += abs(pix1[13] - avg2(pix2[13], pix2[14]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1039 s += abs(pix1[14] - avg2(pix2[14], pix2[15]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1040 s += abs(pix1[15] - avg2(pix2[15], pix2[16]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1041 pix1 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1042 pix2 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1043 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
1044 return s;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1045 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
1046
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1047 int pix_abs16x16_y2_c(UINT8 *pix1, UINT8 *pix2, int line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1048 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
1049 int s, i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1050 UINT8 *pix3 = pix2 + line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1051
986e461dc072 Initial revision
glantau
parents:
diff changeset
1052 s = 0;
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1053 for(i=0;i<16;i++) {
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1054 s += abs(pix1[0] - avg2(pix2[0], pix3[0]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1055 s += abs(pix1[1] - avg2(pix2[1], pix3[1]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1056 s += abs(pix1[2] - avg2(pix2[2], pix3[2]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1057 s += abs(pix1[3] - avg2(pix2[3], pix3[3]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1058 s += abs(pix1[4] - avg2(pix2[4], pix3[4]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1059 s += abs(pix1[5] - avg2(pix2[5], pix3[5]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1060 s += abs(pix1[6] - avg2(pix2[6], pix3[6]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1061 s += abs(pix1[7] - avg2(pix2[7], pix3[7]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1062 s += abs(pix1[8] - avg2(pix2[8], pix3[8]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1063 s += abs(pix1[9] - avg2(pix2[9], pix3[9]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1064 s += abs(pix1[10] - avg2(pix2[10], pix3[10]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1065 s += abs(pix1[11] - avg2(pix2[11], pix3[11]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1066 s += abs(pix1[12] - avg2(pix2[12], pix3[12]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1067 s += abs(pix1[13] - avg2(pix2[13], pix3[13]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1068 s += abs(pix1[14] - avg2(pix2[14], pix3[14]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1069 s += abs(pix1[15] - avg2(pix2[15], pix3[15]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1070 pix1 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1071 pix2 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1072 pix3 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1073 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
1074 return s;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1075 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
1076
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1077 int pix_abs16x16_xy2_c(UINT8 *pix1, UINT8 *pix2, int line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1078 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
1079 int s, i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1080 UINT8 *pix3 = pix2 + line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1081
986e461dc072 Initial revision
glantau
parents:
diff changeset
1082 s = 0;
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1083 for(i=0;i<16;i++) {
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1084 s += abs(pix1[0] - avg4(pix2[0], pix2[1], pix3[0], pix3[1]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1085 s += abs(pix1[1] - avg4(pix2[1], pix2[2], pix3[1], pix3[2]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1086 s += abs(pix1[2] - avg4(pix2[2], pix2[3], pix3[2], pix3[3]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1087 s += abs(pix1[3] - avg4(pix2[3], pix2[4], pix3[3], pix3[4]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1088 s += abs(pix1[4] - avg4(pix2[4], pix2[5], pix3[4], pix3[5]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1089 s += abs(pix1[5] - avg4(pix2[5], pix2[6], pix3[5], pix3[6]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1090 s += abs(pix1[6] - avg4(pix2[6], pix2[7], pix3[6], pix3[7]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1091 s += abs(pix1[7] - avg4(pix2[7], pix2[8], pix3[7], pix3[8]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1092 s += abs(pix1[8] - avg4(pix2[8], pix2[9], pix3[8], pix3[9]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1093 s += abs(pix1[9] - avg4(pix2[9], pix2[10], pix3[9], pix3[10]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1094 s += abs(pix1[10] - avg4(pix2[10], pix2[11], pix3[10], pix3[11]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1095 s += abs(pix1[11] - avg4(pix2[11], pix2[12], pix3[11], pix3[12]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1096 s += abs(pix1[12] - avg4(pix2[12], pix2[13], pix3[12], pix3[13]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1097 s += abs(pix1[13] - avg4(pix2[13], pix2[14], pix3[13], pix3[14]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1098 s += abs(pix1[14] - avg4(pix2[14], pix2[15], pix3[14], pix3[15]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1099 s += abs(pix1[15] - avg4(pix2[15], pix2[16], pix3[15], pix3[16]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1100 pix1 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1101 pix2 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1102 pix3 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1103 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
1104 return s;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1105 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
1106
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1107 int pix_abs8x8_c(UINT8 *pix1, UINT8 *pix2, int line_size)
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1108 {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1109 int s, i;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1110
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1111 s = 0;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1112 for(i=0;i<8;i++) {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1113 s += abs(pix1[0] - pix2[0]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1114 s += abs(pix1[1] - pix2[1]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1115 s += abs(pix1[2] - pix2[2]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1116 s += abs(pix1[3] - pix2[3]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1117 s += abs(pix1[4] - pix2[4]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1118 s += abs(pix1[5] - pix2[5]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1119 s += abs(pix1[6] - pix2[6]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1120 s += abs(pix1[7] - pix2[7]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1121 pix1 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1122 pix2 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1123 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1124 return s;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1125 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1126
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1127 int pix_abs8x8_x2_c(UINT8 *pix1, UINT8 *pix2, int line_size)
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1128 {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1129 int s, i;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1130
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1131 s = 0;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1132 for(i=0;i<8;i++) {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1133 s += abs(pix1[0] - avg2(pix2[0], pix2[1]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1134 s += abs(pix1[1] - avg2(pix2[1], pix2[2]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1135 s += abs(pix1[2] - avg2(pix2[2], pix2[3]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1136 s += abs(pix1[3] - avg2(pix2[3], pix2[4]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1137 s += abs(pix1[4] - avg2(pix2[4], pix2[5]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1138 s += abs(pix1[5] - avg2(pix2[5], pix2[6]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1139 s += abs(pix1[6] - avg2(pix2[6], pix2[7]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1140 s += abs(pix1[7] - avg2(pix2[7], pix2[8]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1141 pix1 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1142 pix2 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1143 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1144 return s;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1145 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1146
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1147 int pix_abs8x8_y2_c(UINT8 *pix1, UINT8 *pix2, int line_size)
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1148 {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1149 int s, i;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1150 UINT8 *pix3 = pix2 + line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1151
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1152 s = 0;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1153 for(i=0;i<8;i++) {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1154 s += abs(pix1[0] - avg2(pix2[0], pix3[0]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1155 s += abs(pix1[1] - avg2(pix2[1], pix3[1]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1156 s += abs(pix1[2] - avg2(pix2[2], pix3[2]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1157 s += abs(pix1[3] - avg2(pix2[3], pix3[3]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1158 s += abs(pix1[4] - avg2(pix2[4], pix3[4]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1159 s += abs(pix1[5] - avg2(pix2[5], pix3[5]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1160 s += abs(pix1[6] - avg2(pix2[6], pix3[6]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1161 s += abs(pix1[7] - avg2(pix2[7], pix3[7]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1162 pix1 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1163 pix2 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1164 pix3 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1165 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1166 return s;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1167 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1168
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1169 int pix_abs8x8_xy2_c(UINT8 *pix1, UINT8 *pix2, int line_size)
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1170 {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1171 int s, i;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1172 UINT8 *pix3 = pix2 + line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1173
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1174 s = 0;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1175 for(i=0;i<8;i++) {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1176 s += abs(pix1[0] - avg4(pix2[0], pix2[1], pix3[0], pix3[1]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1177 s += abs(pix1[1] - avg4(pix2[1], pix2[2], pix3[1], pix3[2]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1178 s += abs(pix1[2] - avg4(pix2[2], pix2[3], pix3[2], pix3[3]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1179 s += abs(pix1[3] - avg4(pix2[3], pix2[4], pix3[3], pix3[4]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1180 s += abs(pix1[4] - avg4(pix2[4], pix2[5], pix3[4], pix3[5]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1181 s += abs(pix1[5] - avg4(pix2[5], pix2[6], pix3[5], pix3[6]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1182 s += abs(pix1[6] - avg4(pix2[6], pix2[7], pix3[6], pix3[7]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1183 s += abs(pix1[7] - avg4(pix2[7], pix2[8], pix3[7], pix3[8]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1184 pix1 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1185 pix2 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1186 pix3 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1187 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1188 return s;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1189 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1190
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1191 /* permute block according so that it corresponds to the MMX idct
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1192 order */
174
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1193 #ifdef SIMPLE_IDCT
190
9e0e56869d05 fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents: 174
diff changeset
1194 /* general permutation, but perhaps slightly slower */
174
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1195 void block_permute(INT16 *block)
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1196 {
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1197 int i;
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1198 INT16 temp[64];
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1199
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1200 for(i=0; i<64; i++) temp[ block_permute_op(i) ] = block[i];
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1201
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1202 for(i=0; i<64; i++) block[i] = temp[i];
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1203 }
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1204 #else
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1205
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1206 void block_permute(INT16 *block)
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1207 {
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1208 int tmp1, tmp2, tmp3, tmp4, tmp5, tmp6;
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1209 int i;
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1210
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1211 for(i=0;i<8;i++) {
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1212 tmp1 = block[1];
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1213 tmp2 = block[2];
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1214 tmp3 = block[3];
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1215 tmp4 = block[4];
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1216 tmp5 = block[5];
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1217 tmp6 = block[6];
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1218 block[1] = tmp2;
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1219 block[2] = tmp4;
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1220 block[3] = tmp6;
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1221 block[4] = tmp1;
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1222 block[5] = tmp3;
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1223 block[6] = tmp5;
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1224 block += 8;
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1225 }
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1226 }
174
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1227 #endif
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1228
296
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
1229 void clear_blocks_c(DCTELEM *blocks)
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
1230 {
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
1231 memset(blocks, 0, sizeof(DCTELEM)*6*64);
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
1232 }
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
1233
480
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1234 /* XXX: those functions should be suppressed ASAP when all IDCTs are
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1235 converted */
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1236 void gen_idct_put(UINT8 *dest, int line_size, DCTELEM *block)
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1237 {
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1238 ff_idct (block);
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1239 put_pixels_clamped(block, dest, line_size);
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1240 }
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1241
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1242 void gen_idct_add(UINT8 *dest, int line_size, DCTELEM *block)
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1243 {
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1244 ff_idct (block);
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1245 add_pixels_clamped(block, dest, line_size);
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1246 }
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1247
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1248 void dsputil_init(void)
986e461dc072 Initial revision
glantau
parents:
diff changeset
1249 {
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1250 int i, j;
88
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1251 int use_permuted_idct;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1252
986e461dc072 Initial revision
glantau
parents:
diff changeset
1253 for(i=0;i<256;i++) cropTbl[i + MAX_NEG_CROP] = i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1254 for(i=0;i<MAX_NEG_CROP;i++) {
986e461dc072 Initial revision
glantau
parents:
diff changeset
1255 cropTbl[i] = 0;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1256 cropTbl[i + MAX_NEG_CROP + 256] = 255;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1257 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
1258
986e461dc072 Initial revision
glantau
parents:
diff changeset
1259 for(i=0;i<512;i++) {
986e461dc072 Initial revision
glantau
parents:
diff changeset
1260 squareTbl[i] = (i - 256) * (i - 256);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1261 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
1262
174
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1263 #ifdef SIMPLE_IDCT
480
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1264 ff_idct = NULL;
174
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1265 #else
19
82d4c9be9873 MMX/MMXEXT iDCT support, using external functions currently defined in libmpeg2
arpi_esp
parents: 6
diff changeset
1266 ff_idct = j_rev_dct;
174
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1267 #endif
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1268 get_pixels = get_pixels_c;
324
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
1269 diff_pixels = diff_pixels_c;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1270 put_pixels_clamped = put_pixels_clamped_c;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1271 add_pixels_clamped = add_pixels_clamped_c;
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1272 gmc1= gmc1_c;
296
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
1273 clear_blocks= clear_blocks_c;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1274
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1275 pix_abs16x16 = pix_abs16x16_c;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1276 pix_abs16x16_x2 = pix_abs16x16_x2_c;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1277 pix_abs16x16_y2 = pix_abs16x16_y2_c;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1278 pix_abs16x16_xy2 = pix_abs16x16_xy2_c;
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1279 pix_abs8x8 = pix_abs8x8_c;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1280 pix_abs8x8_x2 = pix_abs8x8_x2_c;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1281 pix_abs8x8_y2 = pix_abs8x8_y2_c;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1282 pix_abs8x8_xy2 = pix_abs8x8_xy2_c;
474
11dbd00682fc avoid name clash with libjpeg - added missing externs
bellard
parents: 468
diff changeset
1283 av_fdct = fdct_ifast;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1284
88
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1285 use_permuted_idct = 1;
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1286
2
2e2c46c87460 fixed config for direct mplayer build compatibility
glantau
parents: 0
diff changeset
1287 #ifdef HAVE_MMX
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1288 dsputil_init_mmx();
986e461dc072 Initial revision
glantau
parents:
diff changeset
1289 #endif
62
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
1290 #ifdef ARCH_ARMV4L
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
1291 dsputil_init_armv4l();
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
1292 #endif
88
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1293 #ifdef HAVE_MLIB
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1294 dsputil_init_mlib();
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1295 use_permuted_idct = 0;
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1296 #endif
214
73df666cacc7 Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents: 209
diff changeset
1297 #ifdef ARCH_ALPHA
73df666cacc7 Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents: 209
diff changeset
1298 dsputil_init_alpha();
73df666cacc7 Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents: 209
diff changeset
1299 use_permuted_idct = 0;
73df666cacc7 Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents: 209
diff changeset
1300 #endif
88
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1301
174
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1302 #ifdef SIMPLE_IDCT
480
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1303 if (ff_idct == NULL) {
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1304 ff_idct_put = simple_idct_put;
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1305 ff_idct_add = simple_idct_add;
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1306 use_permuted_idct=0;
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1307 } else {
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1308 ff_idct_put = gen_idct_put;
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1309 ff_idct_add = gen_idct_add;
a7cbee351b55 added ff_idct_put/add
bellard
parents: 474
diff changeset
1310 }
174
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1311 #endif
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1312
190
9e0e56869d05 fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents: 174
diff changeset
1313 if(use_permuted_idct)
9e0e56869d05 fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents: 174
diff changeset
1314 #ifdef SIMPLE_IDCT
9e0e56869d05 fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents: 174
diff changeset
1315 for(i=0; i<64; i++) permutation[i]= simple_mmx_permutation[i];
9e0e56869d05 fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents: 174
diff changeset
1316 #else
9e0e56869d05 fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents: 174
diff changeset
1317 for(i=0; i<64; i++) permutation[i]= (i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2);
9e0e56869d05 fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents: 174
diff changeset
1318 #endif
9e0e56869d05 fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents: 174
diff changeset
1319 else
9e0e56869d05 fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents: 174
diff changeset
1320 for(i=0; i<64; i++) permutation[i]=i;
9e0e56869d05 fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents: 174
diff changeset
1321
220
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
1322 for(i=0; i<64; i++) inv_zigzag_direct16[zigzag_direct[i]]= i+1;
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
1323 for(i=0; i<64; i++) zigzag_direct_noperm[i]= zigzag_direct[i];
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
1324
88
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1325 if (use_permuted_idct) {
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1326 /* permute for IDCT */
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1327 for(i=0;i<64;i++) {
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1328 j = zigzag_direct[i];
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1329 zigzag_direct[i] = block_permute_op(j);
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1330 j = ff_alternate_horizontal_scan[i];
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1331 ff_alternate_horizontal_scan[i] = block_permute_op(j);
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1332 j = ff_alternate_vertical_scan[i];
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1333 ff_alternate_vertical_scan[i] = block_permute_op(j);
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1334 }
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1335 block_permute(default_intra_matrix);
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1336 block_permute(default_non_intra_matrix);
312
8cf5507e6ca5 mpeg4 mpeg quantizer support
michaelni
parents: 296
diff changeset
1337 block_permute(ff_mpeg4_default_intra_matrix);
8cf5507e6ca5 mpeg4 mpeg quantizer support
michaelni
parents: 296
diff changeset
1338 block_permute(ff_mpeg4_default_non_intra_matrix);
88
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1339 }
200
6ab301aaa652 (commit by michael)
arpi_esp
parents: 190
diff changeset
1340
6ab301aaa652 (commit by michael)
arpi_esp
parents: 190
diff changeset
1341 build_zigzag_end();
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1342 }
252
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1343
403
2c3e25f4c496 removed unused stuff - added dsputil_set_bit_exact() support for easier testing
glantau
parents: 398
diff changeset
1344 /* remove any non bit exact operation (testing purpose) */
2c3e25f4c496 removed unused stuff - added dsputil_set_bit_exact() support for easier testing
glantau
parents: 398
diff changeset
1345 void avcodec_set_bit_exact(void)
2c3e25f4c496 removed unused stuff - added dsputil_set_bit_exact() support for easier testing
glantau
parents: 398
diff changeset
1346 {
2c3e25f4c496 removed unused stuff - added dsputil_set_bit_exact() support for easier testing
glantau
parents: 398
diff changeset
1347 #ifdef HAVE_MMX
2c3e25f4c496 removed unused stuff - added dsputil_set_bit_exact() support for easier testing
glantau
parents: 398
diff changeset
1348 dsputil_set_bit_exact_mmx();
2c3e25f4c496 removed unused stuff - added dsputil_set_bit_exact() support for easier testing
glantau
parents: 398
diff changeset
1349 #endif
2c3e25f4c496 removed unused stuff - added dsputil_set_bit_exact() support for easier testing
glantau
parents: 398
diff changeset
1350 }
2c3e25f4c496 removed unused stuff - added dsputil_set_bit_exact() support for easier testing
glantau
parents: 398
diff changeset
1351
252
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1352 void get_psnr(UINT8 *orig_image[3], UINT8 *coded_image[3],
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1353 int orig_linesize[3], int coded_linesize,
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1354 AVCodecContext *avctx)
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1355 {
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1356 int quad, diff, x, y;
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1357 UINT8 *orig, *coded;
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1358 UINT32 *sq = squareTbl + 256;
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1359
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1360 quad = 0;
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1361 diff = 0;
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1362
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1363 /* Luminance */
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1364 orig = orig_image[0];
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1365 coded = coded_image[0];
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1366
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1367 for (y=0;y<avctx->height;y++) {
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1368 for (x=0;x<avctx->width;x++) {
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1369 diff = *(orig + x) - *(coded + x);
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1370 quad += sq[diff];
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1371 }
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1372 orig += orig_linesize[0];
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1373 coded += coded_linesize;
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1374 }
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1375
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1376 avctx->psnr_y = (float) quad / (float) (avctx->width * avctx->height);
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1377
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1378 if (avctx->psnr_y) {
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1379 avctx->psnr_y = (float) (255 * 255) / avctx->psnr_y;
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1380 avctx->psnr_y = 10 * (float) log10 (avctx->psnr_y);
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1381 } else
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1382 avctx->psnr_y = 99.99;
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1383 }
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1384