annotate dsputil.c @ 844:1bd52b59bcd1 libavcodec

fixing -sameq with mpeg2
author michaelni
date Sat, 09 Nov 2002 13:57:54 +0000
parents d4726182dfd2
children eacc2dd8fd9d
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"
986e461dc072 Initial revision
glantau
parents:
diff changeset
23
986e461dc072 Initial revision
glantau
parents:
diff changeset
24 void (*get_pixels)(DCTELEM *block, const UINT8 *pixels, int line_size);
324
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
25 void (*diff_pixels)(DCTELEM *block, const UINT8 *s1, const UINT8 *s2, int stride);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
26 void (*put_pixels_clamped)(const DCTELEM *block, UINT8 *pixels, int line_size);
986e461dc072 Initial revision
glantau
parents:
diff changeset
27 void (*add_pixels_clamped)(const DCTELEM *block, UINT8 *pixels, int line_size);
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
28 void (*ff_gmc1)(UINT8 *dst, UINT8 *src, int srcStride, int h, int x16, int y16, int rounder);
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
29 void (*ff_gmc )(UINT8 *dst, UINT8 *src, int stride, int h, int ox, int oy,
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
30 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
296
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
31 void (*clear_blocks)(DCTELEM *blocks);
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
32 int (*pix_sum)(UINT8 * pix, int line_size);
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
33 int (*pix_norm1)(UINT8 * pix, int line_size);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
34
986e461dc072 Initial revision
glantau
parents:
diff changeset
35 op_pixels_abs_func pix_abs16x16;
986e461dc072 Initial revision
glantau
parents:
diff changeset
36 op_pixels_abs_func pix_abs16x16_x2;
986e461dc072 Initial revision
glantau
parents:
diff changeset
37 op_pixels_abs_func pix_abs16x16_y2;
986e461dc072 Initial revision
glantau
parents:
diff changeset
38 op_pixels_abs_func pix_abs16x16_xy2;
986e461dc072 Initial revision
glantau
parents:
diff changeset
39
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
40 op_pixels_abs_func pix_abs8x8;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
41 op_pixels_abs_func pix_abs8x8_x2;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
42 op_pixels_abs_func pix_abs8x8_y2;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
43 op_pixels_abs_func pix_abs8x8_xy2;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
44
676
c3bdb00a98a9 dont store version for bit-exact tests
michaelni
parents: 651
diff changeset
45 int ff_bit_exact=0;
c3bdb00a98a9 dont store version for bit-exact tests
michaelni
parents: 651
diff changeset
46
50
4ea4c10d03d8 export for imgconvert usage
glantau
parents: 34
diff changeset
47 UINT8 cropTbl[256 + 2 * MAX_NEG_CROP];
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
48 UINT32 squareTbl[512];
986e461dc072 Initial revision
glantau
parents:
diff changeset
49
706
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 689
diff changeset
50 const UINT8 ff_zigzag_direct[64] = {
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 689
diff changeset
51 0, 1, 8, 16, 9, 2, 3, 10,
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 689
diff changeset
52 17, 24, 32, 25, 18, 11, 4, 5,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
53 12, 19, 26, 33, 40, 48, 41, 34,
706
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 689
diff changeset
54 27, 20, 13, 6, 7, 14, 21, 28,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
55 35, 42, 49, 56, 57, 50, 43, 36,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
56 29, 22, 15, 23, 30, 37, 44, 51,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
57 58, 59, 52, 45, 38, 31, 39, 46,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
58 53, 60, 61, 54, 47, 55, 62, 63
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
59 };
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
60
220
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
61 /* not permutated inverse zigzag_direct + 1 for MMX quantizer */
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
62 UINT16 __align8 inv_zigzag_direct16[64];
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
63
706
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 689
diff changeset
64 const UINT8 ff_alternate_horizontal_scan[64] = {
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 689
diff changeset
65 0, 1, 2, 3, 8, 9, 16, 17,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
66 10, 11, 4, 5, 6, 7, 15, 14,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
67 13, 12, 19, 18, 24, 25, 32, 33,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
68 26, 27, 20, 21, 22, 23, 28, 29,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
69 30, 31, 34, 35, 40, 41, 48, 49,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
70 42, 43, 36, 37, 38, 39, 44, 45,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
71 46, 47, 50, 51, 56, 57, 58, 59,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
72 52, 53, 54, 55, 60, 61, 62, 63,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
73 };
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
74
706
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 689
diff changeset
75 const UINT8 ff_alternate_vertical_scan[64] = {
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 689
diff changeset
76 0, 8, 16, 24, 1, 9, 2, 10,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
77 17, 25, 32, 40, 48, 56, 57, 49,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
78 41, 33, 26, 18, 3, 11, 4, 12,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
79 19, 27, 34, 42, 50, 58, 35, 43,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
80 51, 59, 20, 28, 5, 13, 6, 14,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
81 21, 29, 36, 44, 52, 60, 37, 45,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
82 53, 61, 22, 30, 7, 15, 23, 31,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
83 38, 46, 54, 62, 39, 47, 55, 63,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
84 };
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
85
220
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
86 /* a*inverse[b]>>32 == a/b for all 0<=a<=65536 && 2<=b<=255 */
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
87 UINT32 inverse[256]={
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
88 0, 4294967295U,2147483648U,1431655766, 1073741824, 858993460, 715827883, 613566757,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
89 536870912, 477218589, 429496730, 390451573, 357913942, 330382100, 306783379, 286331154,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
90 268435456, 252645136, 238609295, 226050911, 214748365, 204522253, 195225787, 186737709,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
91 178956971, 171798692, 165191050, 159072863, 153391690, 148102321, 143165577, 138547333,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
92 134217728, 130150525, 126322568, 122713352, 119304648, 116080198, 113025456, 110127367,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
93 107374183, 104755300, 102261127, 99882961, 97612894, 95443718, 93368855, 91382283,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
94 89478486, 87652394, 85899346, 84215046, 82595525, 81037119, 79536432, 78090315,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
95 76695845, 75350304, 74051161, 72796056, 71582789, 70409300, 69273667, 68174085,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
96 67108864, 66076420, 65075263, 64103990, 63161284, 62245903, 61356676, 60492498,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
97 59652324, 58835169, 58040099, 57266231, 56512728, 55778797, 55063684, 54366675,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
98 53687092, 53024288, 52377650, 51746594, 51130564, 50529028, 49941481, 49367441,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
99 48806447, 48258060, 47721859, 47197443, 46684428, 46182445, 45691142, 45210183,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
100 44739243, 44278014, 43826197, 43383509, 42949673, 42524429, 42107523, 41698712,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
101 41297763, 40904451, 40518560, 40139882, 39768216, 39403370, 39045158, 38693400,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
102 38347923, 38008561, 37675152, 37347542, 37025581, 36709123, 36398028, 36092163,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
103 35791395, 35495598, 35204650, 34918434, 34636834, 34359739, 34087043, 33818641,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
104 33554432, 33294321, 33038210, 32786010, 32537632, 32292988, 32051995, 31814573,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
105 31580642, 31350127, 31122952, 30899046, 30678338, 30460761, 30246249, 30034737,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
106 29826162, 29620465, 29417585, 29217465, 29020050, 28825284, 28633116, 28443493,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
107 28256364, 28071682, 27889399, 27709467, 27531842, 27356480, 27183338, 27012373,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
108 26843546, 26676816, 26512144, 26349493, 26188825, 26030105, 25873297, 25718368,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
109 25565282, 25414008, 25264514, 25116768, 24970741, 24826401, 24683721, 24542671,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
110 24403224, 24265352, 24129030, 23994231, 23860930, 23729102, 23598722, 23469767,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
111 23342214, 23216040, 23091223, 22967740, 22845571, 22724695, 22605092, 22486740,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
112 22369622, 22253717, 22139007, 22025474, 21913099, 21801865, 21691755, 21582751,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
113 21474837, 21367997, 21262215, 21157475, 21053762, 20951060, 20849356, 20748635,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
114 20648882, 20550083, 20452226, 20355296, 20259280, 20164166, 20069941, 19976593,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
115 19884108, 19792477, 19701685, 19611723, 19522579, 19434242, 19346700, 19259944,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
116 19173962, 19088744, 19004281, 18920561, 18837576, 18755316, 18673771, 18592933,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
117 18512791, 18433337, 18354562, 18276457, 18199014, 18122225, 18046082, 17970575,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
118 17895698, 17821442, 17747799, 17674763, 17602325, 17530479, 17459217, 17388532,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
119 17318417, 17248865, 17179870, 17111424, 17043522, 16976156, 16909321, 16843010,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
120 };
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
121
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
122 int pix_sum_c(UINT8 * pix, int line_size)
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
123 {
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
124 int s, i, j;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
125
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
126 s = 0;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
127 for (i = 0; i < 16; i++) {
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
128 for (j = 0; j < 16; j += 8) {
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
129 s += pix[0];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
130 s += pix[1];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
131 s += pix[2];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
132 s += pix[3];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
133 s += pix[4];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
134 s += pix[5];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
135 s += pix[6];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
136 s += pix[7];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
137 pix += 8;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
138 }
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
139 pix += line_size - 16;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
140 }
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
141 return s;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
142 }
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
143
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
144 int pix_norm1_c(UINT8 * pix, int line_size)
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
145 {
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
146 int s, i, j;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
147 UINT32 *sq = squareTbl + 256;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
148
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
149 s = 0;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
150 for (i = 0; i < 16; i++) {
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
151 for (j = 0; j < 16; j += 8) {
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
152 s += sq[pix[0]];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
153 s += sq[pix[1]];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
154 s += sq[pix[2]];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
155 s += sq[pix[3]];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
156 s += sq[pix[4]];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
157 s += sq[pix[5]];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
158 s += sq[pix[6]];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
159 s += sq[pix[7]];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
160 pix += 8;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
161 }
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
162 pix += line_size - 16;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
163 }
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
164 return s;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
165 }
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
166
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
167
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
168 void get_pixels_c(DCTELEM *restrict block, const UINT8 *pixels, int line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
169 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
170 int i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
171
986e461dc072 Initial revision
glantau
parents:
diff changeset
172 /* read the pixels */
986e461dc072 Initial revision
glantau
parents:
diff changeset
173 for(i=0;i<8;i++) {
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
174 block[0] = pixels[0];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
175 block[1] = pixels[1];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
176 block[2] = pixels[2];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
177 block[3] = pixels[3];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
178 block[4] = pixels[4];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
179 block[5] = pixels[5];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
180 block[6] = pixels[6];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
181 block[7] = pixels[7];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
182 pixels += line_size;
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
183 block += 8;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
184 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
185 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
186
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
187 void diff_pixels_c(DCTELEM *restrict block, const UINT8 *s1, const UINT8 *s2,
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
188 int stride){
324
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
189 int i;
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
190
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
191 /* read the pixels */
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
192 for(i=0;i<8;i++) {
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
193 block[0] = s1[0] - s2[0];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
194 block[1] = s1[1] - s2[1];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
195 block[2] = s1[2] - s2[2];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
196 block[3] = s1[3] - s2[3];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
197 block[4] = s1[4] - s2[4];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
198 block[5] = s1[5] - s2[5];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
199 block[6] = s1[6] - s2[6];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
200 block[7] = s1[7] - s2[7];
324
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
201 s1 += stride;
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
202 s2 += stride;
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
203 block += 8;
324
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
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
207
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
208 void put_pixels_clamped_c(const DCTELEM *block, UINT8 *restrict pixels,
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
209 int line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
210 {
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 for(i=0;i<8;i++) {
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
216 pixels[0] = cm[block[0]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
217 pixels[1] = cm[block[1]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
218 pixels[2] = cm[block[2]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
219 pixels[3] = cm[block[3]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
220 pixels[4] = cm[block[4]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
221 pixels[5] = cm[block[5]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
222 pixels[6] = cm[block[6]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
223 pixels[7] = cm[block[7]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
224
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
225 pixels += line_size;
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
226 block += 8;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
227 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
228 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
229
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
230 void add_pixels_clamped_c(const DCTELEM *block, UINT8 *restrict pixels,
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
231 int line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
232 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
233 int i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
234 UINT8 *cm = cropTbl + MAX_NEG_CROP;
986e461dc072 Initial revision
glantau
parents:
diff changeset
235
986e461dc072 Initial revision
glantau
parents:
diff changeset
236 /* read the pixels */
986e461dc072 Initial revision
glantau
parents:
diff changeset
237 for(i=0;i<8;i++) {
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
238 pixels[0] = cm[pixels[0] + block[0]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
239 pixels[1] = cm[pixels[1] + block[1]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
240 pixels[2] = cm[pixels[2] + block[2]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
241 pixels[3] = cm[pixels[3] + block[3]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
242 pixels[4] = cm[pixels[4] + block[4]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
243 pixels[5] = cm[pixels[5] + block[5]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
244 pixels[6] = cm[pixels[6] + block[6]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
245 pixels[7] = cm[pixels[7] + block[7]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
246 pixels += line_size;
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
247 block += 8;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
248 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
249 }
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
250 #if 0
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
251
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
252 #define PIXOP2(OPNAME, OP) \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
253 static void OPNAME ## _pixels(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
254 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
255 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
256 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
257 OP(*((uint64_t*)block), LD64(pixels));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
258 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
259 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
260 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
261 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
262 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
263 static void OPNAME ## _no_rnd_pixels_x2(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
264 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
265 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
266 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
267 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
268 const uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
269 OP(*((uint64_t*)block), (a&b) + (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
270 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
271 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
272 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
273 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
274 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
275 static void OPNAME ## _pixels_x2(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
276 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
277 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
278 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
279 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
280 const uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
281 OP(*((uint64_t*)block), (a|b) - (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
282 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
283 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
284 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
285 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
286 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
287 static void OPNAME ## _no_rnd_pixels_y2(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
288 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
289 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
290 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
291 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
292 const uint64_t b= LD64(pixels+line_size);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
293 OP(*((uint64_t*)block), (a&b) + (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
294 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
295 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
296 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
297 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
298 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
299 static void OPNAME ## _pixels_y2(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
300 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
301 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
302 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
303 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
304 const uint64_t b= LD64(pixels+line_size);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
305 OP(*((uint64_t*)block), (a|b) - (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
306 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
307 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
308 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
309 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
310 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
311 static void OPNAME ## _pixels_xy2(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
312 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
313 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
314 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
315 const uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
316 uint64_t l0= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
317 + (b&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
318 + 0x0202020202020202ULL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
319 uint64_t h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
320 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
321 uint64_t l1,h1;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
322 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
323 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
324 for(i=0; i<h; i+=2){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
325 uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
326 uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
327 l1= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
328 + (b&0x0303030303030303ULL);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
329 h1= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
330 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
331 OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
332 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
333 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
334 a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
335 b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
336 l0= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
337 + (b&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
338 + 0x0202020202020202ULL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
339 h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
340 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
341 OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
342 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
343 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
344 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
345 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
346 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
347 static void OPNAME ## _no_rnd_pixels_xy2(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
348 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
349 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
350 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
351 const uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
352 uint64_t l0= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
353 + (b&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
354 + 0x0101010101010101ULL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
355 uint64_t h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
356 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
357 uint64_t l1,h1;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
358 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
359 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
360 for(i=0; i<h; i+=2){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
361 uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
362 uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
363 l1= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
364 + (b&0x0303030303030303ULL);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
365 h1= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
366 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
367 OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
368 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
369 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
370 a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
371 b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
372 l0= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
373 + (b&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
374 + 0x0101010101010101ULL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
375 h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
376 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
377 OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
378 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
379 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
380 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
381 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
382 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
383 CALL_2X_PIXELS(OPNAME ## _pixels16 , OPNAME ## _pixels , 8)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
384 CALL_2X_PIXELS(OPNAME ## _pixels16_x2 , OPNAME ## _pixels_x2 , 8)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
385 CALL_2X_PIXELS(OPNAME ## _pixels16_y2 , OPNAME ## _pixels_y2 , 8)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
386 CALL_2X_PIXELS(OPNAME ## _pixels16_xy2, OPNAME ## _pixels_xy2, 8)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
387 CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_x2 , OPNAME ## _no_rnd_pixels_x2 , 8)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
388 CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_y2 , OPNAME ## _no_rnd_pixels_y2 , 8)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
389 CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_xy2, OPNAME ## _no_rnd_pixels_xy2, 8)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
390 \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
391 void (*OPNAME ## _pixels_tab[2][4])(uint8_t *block, const uint8_t *pixels, int line_size, int h) = {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
392 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
393 OPNAME ## _pixels,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
394 OPNAME ## _pixels_x2,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
395 OPNAME ## _pixels_y2,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
396 OPNAME ## _pixels_xy2},\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
397 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
398 OPNAME ## _pixels16,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
399 OPNAME ## _pixels16_x2,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
400 OPNAME ## _pixels16_y2,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
401 OPNAME ## _pixels16_xy2}\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
402 };\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
403 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
404 void (*OPNAME ## _no_rnd_pixels_tab[2][4])(uint8_t *block, const uint8_t *pixels, int line_size, int h) = {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
405 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
406 OPNAME ## _pixels,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
407 OPNAME ## _no_rnd_pixels_x2,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
408 OPNAME ## _no_rnd_pixels_y2,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
409 OPNAME ## _no_rnd_pixels_xy2},\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
410 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
411 OPNAME ## _pixels16,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
412 OPNAME ## _no_rnd_pixels16_x2,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
413 OPNAME ## _no_rnd_pixels16_y2,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
414 OPNAME ## _no_rnd_pixels16_xy2}\
385
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 #define op_avg(a, b) a = ( ((a)|(b)) - ((((a)^(b))&0xFEFEFEFEFEFEFEFEULL)>>1) )
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
418 #else // 64 bit variant
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
419
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
420 #define PIXOP2(OPNAME, OP) \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
421 static void OPNAME ## _pixels8(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
422 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
423 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
424 OP(*((uint32_t*)(block )), LD32(pixels ));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
425 OP(*((uint32_t*)(block+4)), LD32(pixels+4));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
426 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
427 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
428 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
429 }\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
430 static inline void OPNAME ## _no_rnd_pixels8(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
431 OPNAME ## _pixels8(block, pixels, line_size, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
432 }\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
433 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
434 static inline void OPNAME ## _no_rnd_pixels8_l2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
435 int src_stride1, int src_stride2, int h){\
385
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++){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
438 uint32_t a,b;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
439 a= LD32(&src1[i*src_stride1 ]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
440 b= LD32(&src2[i*src_stride2 ]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
441 OP(*((uint32_t*)&dst[i*dst_stride ]), (a&b) + (((a^b)&0xFEFEFEFEUL)>>1));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
442 a= LD32(&src1[i*src_stride1+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
443 b= LD32(&src2[i*src_stride2+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
444 OP(*((uint32_t*)&dst[i*dst_stride+4]), (a&b) + (((a^b)&0xFEFEFEFEUL)>>1));\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
445 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
446 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
447 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
448 static inline void OPNAME ## _pixels8_l2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
449 int src_stride1, int src_stride2, int h){\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
450 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
451 for(i=0; i<h; i++){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
452 uint32_t a,b;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
453 a= LD32(&src1[i*src_stride1 ]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
454 b= LD32(&src2[i*src_stride2 ]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
455 OP(*((uint32_t*)&dst[i*dst_stride ]), (a|b) - (((a^b)&0xFEFEFEFEUL)>>1));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
456 a= LD32(&src1[i*src_stride1+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
457 b= LD32(&src2[i*src_stride2+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
458 OP(*((uint32_t*)&dst[i*dst_stride+4]), (a|b) - (((a^b)&0xFEFEFEFEUL)>>1));\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
459 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
460 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
461 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
462 static inline void OPNAME ## _pixels16_l2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
463 int src_stride1, int src_stride2, int h){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
464 OPNAME ## _pixels8_l2(dst , src1 , src2 , dst_stride, src_stride1, src_stride2, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
465 OPNAME ## _pixels8_l2(dst+8, src1+8, src2+8, dst_stride, src_stride1, src_stride2, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
466 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
467 \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
468 static inline void OPNAME ## _no_rnd_pixels16_l2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
469 int src_stride1, int src_stride2, int h){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
470 OPNAME ## _no_rnd_pixels8_l2(dst , src1 , src2 , dst_stride, src_stride1, src_stride2, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
471 OPNAME ## _no_rnd_pixels8_l2(dst+8, src1+8, src2+8, dst_stride, src_stride1, src_stride2, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
472 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
473 \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
474 static inline void OPNAME ## _no_rnd_pixels8_x2(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
475 OPNAME ## _no_rnd_pixels8_l2(block, pixels, pixels+1, line_size, line_size, line_size, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
476 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
477 \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
478 static inline void OPNAME ## _pixels8_x2(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
479 OPNAME ## _pixels8_l2(block, pixels, pixels+1, line_size, line_size, line_size, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
480 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
481 \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
482 static inline void OPNAME ## _no_rnd_pixels8_y2(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
483 OPNAME ## _no_rnd_pixels8_l2(block, pixels, pixels+line_size, line_size, line_size, line_size, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
484 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
485 \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
486 static inline void OPNAME ## _pixels8_y2(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
487 OPNAME ## _pixels8_l2(block, pixels, pixels+line_size, line_size, line_size, line_size, h);\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
488 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
489 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
490 static inline void OPNAME ## _pixels8_l4(uint8_t *dst, const uint8_t *src1, uint8_t *src2, uint8_t *src3, uint8_t *src4,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
491 int dst_stride, int src_stride1, int src_stride2,int src_stride3,int src_stride4, int h){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
492 int i;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
493 for(i=0; i<h; i++){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
494 uint32_t a, b, c, d, l0, l1, h0, h1;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
495 a= LD32(&src1[i*src_stride1]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
496 b= LD32(&src2[i*src_stride2]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
497 c= LD32(&src3[i*src_stride3]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
498 d= LD32(&src4[i*src_stride4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
499 l0= (a&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
500 + (b&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
501 + 0x02020202UL;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
502 h0= ((a&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
503 + ((b&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
504 l1= (c&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
505 + (d&0x03030303UL);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
506 h1= ((c&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
507 + ((d&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
508 OP(*((uint32_t*)&dst[i*dst_stride]), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
509 a= LD32(&src1[i*src_stride1+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
510 b= LD32(&src2[i*src_stride2+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
511 c= LD32(&src3[i*src_stride3+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
512 d= LD32(&src4[i*src_stride4+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
513 l0= (a&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
514 + (b&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
515 + 0x02020202UL;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
516 h0= ((a&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
517 + ((b&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
518 l1= (c&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
519 + (d&0x03030303UL);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
520 h1= ((c&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
521 + ((d&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
522 OP(*((uint32_t*)&dst[i*dst_stride+4]), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
523 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
524 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
525 static inline void OPNAME ## _no_rnd_pixels8_l4(uint8_t *dst, const uint8_t *src1, uint8_t *src2, uint8_t *src3, uint8_t *src4,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
526 int dst_stride, int src_stride1, int src_stride2,int src_stride3,int src_stride4, int h){\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
527 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
528 for(i=0; i<h; i++){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
529 uint32_t a, b, c, d, l0, l1, h0, h1;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
530 a= LD32(&src1[i*src_stride1]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
531 b= LD32(&src2[i*src_stride2]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
532 c= LD32(&src3[i*src_stride3]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
533 d= LD32(&src4[i*src_stride4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
534 l0= (a&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
535 + (b&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
536 + 0x01010101UL;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
537 h0= ((a&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
538 + ((b&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
539 l1= (c&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
540 + (d&0x03030303UL);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
541 h1= ((c&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
542 + ((d&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
543 OP(*((uint32_t*)&dst[i*dst_stride]), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
544 a= LD32(&src1[i*src_stride1+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
545 b= LD32(&src2[i*src_stride2+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
546 c= LD32(&src3[i*src_stride3+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
547 d= LD32(&src4[i*src_stride4+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
548 l0= (a&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
549 + (b&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
550 + 0x01010101UL;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
551 h0= ((a&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
552 + ((b&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
553 l1= (c&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
554 + (d&0x03030303UL);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
555 h1= ((c&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
556 + ((d&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
557 OP(*((uint32_t*)&dst[i*dst_stride+4]), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
558 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
559 }\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
560 static inline void OPNAME ## _pixels16_l4(uint8_t *dst, const uint8_t *src1, uint8_t *src2, uint8_t *src3, uint8_t *src4,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
561 int dst_stride, int src_stride1, int src_stride2,int src_stride3,int src_stride4, int h){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
562 OPNAME ## _pixels8_l4(dst , src1 , src2 , src3 , src4 , dst_stride, src_stride1, src_stride2, src_stride3, src_stride4, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
563 OPNAME ## _pixels8_l4(dst+8, src1+8, src2+8, src3+8, src4+8, dst_stride, src_stride1, src_stride2, src_stride3, src_stride4, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
564 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
565 static inline void OPNAME ## _no_rnd_pixels16_l4(uint8_t *dst, const uint8_t *src1, uint8_t *src2, uint8_t *src3, uint8_t *src4,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
566 int dst_stride, int src_stride1, int src_stride2,int src_stride3,int src_stride4, int h){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
567 OPNAME ## _no_rnd_pixels8_l4(dst , src1 , src2 , src3 , src4 , dst_stride, src_stride1, src_stride2, src_stride3, src_stride4, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
568 OPNAME ## _no_rnd_pixels8_l4(dst+8, src1+8, src2+8, src3+8, src4+8, dst_stride, src_stride1, src_stride2, src_stride3, src_stride4, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
569 }\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
570 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
571 static inline void OPNAME ## _pixels8_xy2(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
572 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
573 int j;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
574 for(j=0; j<2; j++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
575 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
576 const uint32_t a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
577 const uint32_t b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
578 uint32_t l0= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
579 + (b&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
580 + 0x02020202UL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
581 uint32_t h0= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
582 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
583 uint32_t l1,h1;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
584 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
585 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
586 for(i=0; i<h; i+=2){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
587 uint32_t a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
588 uint32_t b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
589 l1= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
590 + (b&0x03030303UL);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
591 h1= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
592 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
593 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
594 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
595 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
596 a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
597 b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
598 l0= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
599 + (b&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
600 + 0x02020202UL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
601 h0= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
602 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
603 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
604 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
605 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
606 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
607 pixels+=4-line_size*(h+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
608 block +=4-line_size*h;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
609 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
610 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
611 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
612 static inline void OPNAME ## _no_rnd_pixels8_xy2(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
613 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
614 int j;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
615 for(j=0; j<2; j++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
616 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
617 const uint32_t a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
618 const uint32_t b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
619 uint32_t l0= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
620 + (b&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
621 + 0x01010101UL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
622 uint32_t h0= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
623 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
624 uint32_t l1,h1;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
625 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
626 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
627 for(i=0; i<h; i+=2){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
628 uint32_t a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
629 uint32_t b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
630 l1= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
631 + (b&0x03030303UL);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
632 h1= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
633 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
634 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
635 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
636 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
637 a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
638 b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
639 l0= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
640 + (b&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
641 + 0x01010101UL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
642 h0= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
643 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
644 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
645 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
646 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
647 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
648 pixels+=4-line_size*(h+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
649 block +=4-line_size*h;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
650 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
651 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
652 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
653 CALL_2X_PIXELS(OPNAME ## _pixels16 , OPNAME ## _pixels8 , 8)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
654 CALL_2X_PIXELS(OPNAME ## _pixels16_x2 , OPNAME ## _pixels8_x2 , 8)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
655 CALL_2X_PIXELS(OPNAME ## _pixels16_y2 , OPNAME ## _pixels8_y2 , 8)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
656 CALL_2X_PIXELS(OPNAME ## _pixels16_xy2, OPNAME ## _pixels8_xy2, 8)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
657 CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16 , OPNAME ## _pixels8 , 8)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
658 CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_x2 , OPNAME ## _no_rnd_pixels8_x2 , 8)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
659 CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_y2 , OPNAME ## _no_rnd_pixels8_y2 , 8)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
660 CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_xy2, OPNAME ## _no_rnd_pixels8_xy2, 8)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
661 \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
662 void (*OPNAME ## _pixels_tab[2][4])(uint8_t *block, const uint8_t *pixels, int line_size, int h) = {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
663 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
664 OPNAME ## _pixels16,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
665 OPNAME ## _pixels16_x2,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
666 OPNAME ## _pixels16_y2,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
667 OPNAME ## _pixels16_xy2},\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
668 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
669 OPNAME ## _pixels8,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
670 OPNAME ## _pixels8_x2,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
671 OPNAME ## _pixels8_y2,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
672 OPNAME ## _pixels8_xy2},\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
673 };\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
674 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
675 void (*OPNAME ## _no_rnd_pixels_tab[2][4])(uint8_t *block, const uint8_t *pixels, int line_size, int h) = {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
676 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
677 OPNAME ## _pixels16,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
678 OPNAME ## _no_rnd_pixels16_x2,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
679 OPNAME ## _no_rnd_pixels16_y2,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
680 OPNAME ## _no_rnd_pixels16_xy2},\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
681 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
682 OPNAME ## _pixels8,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
683 OPNAME ## _no_rnd_pixels8_x2,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
684 OPNAME ## _no_rnd_pixels8_y2,\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
685 OPNAME ## _no_rnd_pixels8_xy2},\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
686 };
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
687
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
688 #define op_avg(a, b) a = ( ((a)|(b)) - ((((a)^(b))&0xFEFEFEFEUL)>>1) )
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
689 #endif
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
690 #define op_put(a, b) a = b
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
691
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
692 PIXOP2(avg, op_avg)
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
693 PIXOP2(put, op_put)
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
694 #undef op_avg
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
695 #undef op_put
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
696
403
2c3e25f4c496 removed unused stuff - added dsputil_set_bit_exact() support for easier testing
glantau
parents: 398
diff changeset
697 #if 0
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
698 /* FIXME this stuff could be removed as its ot really used anymore */
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
699 #define PIXOP(BTYPE, OPNAME, OP, INCR) \
986e461dc072 Initial revision
glantau
parents:
diff changeset
700 \
986e461dc072 Initial revision
glantau
parents:
diff changeset
701 static void OPNAME ## _pixels(BTYPE *block, const UINT8 *pixels, int line_size, int h) \
986e461dc072 Initial revision
glantau
parents:
diff changeset
702 { \
986e461dc072 Initial revision
glantau
parents:
diff changeset
703 BTYPE *p; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
704 const UINT8 *pix; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
705 \
986e461dc072 Initial revision
glantau
parents:
diff changeset
706 p = block; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
707 pix = pixels; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
708 do { \
986e461dc072 Initial revision
glantau
parents:
diff changeset
709 OP(p[0], pix[0]); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
710 OP(p[1], pix[1]); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
711 OP(p[2], pix[2]); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
712 OP(p[3], pix[3]); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
713 OP(p[4], pix[4]); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
714 OP(p[5], pix[5]); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
715 OP(p[6], pix[6]); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
716 OP(p[7], pix[7]); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
717 pix += line_size; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
718 p += INCR; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
719 } while (--h);; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
720 } \
986e461dc072 Initial revision
glantau
parents:
diff changeset
721 \
986e461dc072 Initial revision
glantau
parents:
diff changeset
722 static void OPNAME ## _pixels_x2(BTYPE *block, const UINT8 *pixels, int line_size, int h) \
986e461dc072 Initial revision
glantau
parents:
diff changeset
723 { \
986e461dc072 Initial revision
glantau
parents:
diff changeset
724 BTYPE *p; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
725 const UINT8 *pix; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
726 \
986e461dc072 Initial revision
glantau
parents:
diff changeset
727 p = block; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
728 pix = pixels; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
729 do { \
986e461dc072 Initial revision
glantau
parents:
diff changeset
730 OP(p[0], avg2(pix[0], pix[1])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
731 OP(p[1], avg2(pix[1], pix[2])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
732 OP(p[2], avg2(pix[2], pix[3])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
733 OP(p[3], avg2(pix[3], pix[4])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
734 OP(p[4], avg2(pix[4], pix[5])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
735 OP(p[5], avg2(pix[5], pix[6])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
736 OP(p[6], avg2(pix[6], pix[7])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
737 OP(p[7], avg2(pix[7], pix[8])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
738 pix += line_size; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
739 p += INCR; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
740 } while (--h); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
741 } \
986e461dc072 Initial revision
glantau
parents:
diff changeset
742 \
986e461dc072 Initial revision
glantau
parents:
diff changeset
743 static void OPNAME ## _pixels_y2(BTYPE *block, const UINT8 *pixels, int line_size, int h) \
986e461dc072 Initial revision
glantau
parents:
diff changeset
744 { \
986e461dc072 Initial revision
glantau
parents:
diff changeset
745 BTYPE *p; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
746 const UINT8 *pix; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
747 const UINT8 *pix1; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
748 \
986e461dc072 Initial revision
glantau
parents:
diff changeset
749 p = block; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
750 pix = pixels; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
751 pix1 = pixels + line_size; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
752 do { \
986e461dc072 Initial revision
glantau
parents:
diff changeset
753 OP(p[0], avg2(pix[0], pix1[0])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
754 OP(p[1], avg2(pix[1], pix1[1])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
755 OP(p[2], avg2(pix[2], pix1[2])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
756 OP(p[3], avg2(pix[3], pix1[3])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
757 OP(p[4], avg2(pix[4], pix1[4])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
758 OP(p[5], avg2(pix[5], pix1[5])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
759 OP(p[6], avg2(pix[6], pix1[6])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
760 OP(p[7], avg2(pix[7], pix1[7])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
761 pix += line_size; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
762 pix1 += line_size; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
763 p += INCR; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
764 } while(--h); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
765 } \
986e461dc072 Initial revision
glantau
parents:
diff changeset
766 \
986e461dc072 Initial revision
glantau
parents:
diff changeset
767 static void OPNAME ## _pixels_xy2(BTYPE *block, const UINT8 *pixels, int line_size, int h) \
986e461dc072 Initial revision
glantau
parents:
diff changeset
768 { \
986e461dc072 Initial revision
glantau
parents:
diff changeset
769 BTYPE *p; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
770 const UINT8 *pix; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
771 const UINT8 *pix1; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
772 \
986e461dc072 Initial revision
glantau
parents:
diff changeset
773 p = block; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
774 pix = pixels; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
775 pix1 = pixels + line_size; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
776 do { \
986e461dc072 Initial revision
glantau
parents:
diff changeset
777 OP(p[0], avg4(pix[0], pix[1], pix1[0], pix1[1])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
778 OP(p[1], avg4(pix[1], pix[2], pix1[1], pix1[2])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
779 OP(p[2], avg4(pix[2], pix[3], pix1[2], pix1[3])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
780 OP(p[3], avg4(pix[3], pix[4], pix1[3], pix1[4])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
781 OP(p[4], avg4(pix[4], pix[5], pix1[4], pix1[5])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
782 OP(p[5], avg4(pix[5], pix[6], pix1[5], pix1[6])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
783 OP(p[6], avg4(pix[6], pix[7], pix1[6], pix1[7])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
784 OP(p[7], avg4(pix[7], pix[8], pix1[7], pix1[8])); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
785 pix += line_size; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
786 pix1 += line_size; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
787 p += INCR; \
986e461dc072 Initial revision
glantau
parents:
diff changeset
788 } while(--h); \
986e461dc072 Initial revision
glantau
parents:
diff changeset
789 } \
986e461dc072 Initial revision
glantau
parents:
diff changeset
790 \
986e461dc072 Initial revision
glantau
parents:
diff changeset
791 void (*OPNAME ## _pixels_tab[4])(BTYPE *block, const UINT8 *pixels, int line_size, int h) = { \
986e461dc072 Initial revision
glantau
parents:
diff changeset
792 OPNAME ## _pixels, \
986e461dc072 Initial revision
glantau
parents:
diff changeset
793 OPNAME ## _pixels_x2, \
986e461dc072 Initial revision
glantau
parents:
diff changeset
794 OPNAME ## _pixels_y2, \
986e461dc072 Initial revision
glantau
parents:
diff changeset
795 OPNAME ## _pixels_xy2, \
986e461dc072 Initial revision
glantau
parents:
diff changeset
796 };
986e461dc072 Initial revision
glantau
parents:
diff changeset
797
986e461dc072 Initial revision
glantau
parents:
diff changeset
798 /* rounding primitives */
986e461dc072 Initial revision
glantau
parents:
diff changeset
799 #define avg2(a,b) ((a+b+1)>>1)
986e461dc072 Initial revision
glantau
parents:
diff changeset
800 #define avg4(a,b,c,d) ((a+b+c+d+2)>>2)
986e461dc072 Initial revision
glantau
parents:
diff changeset
801
986e461dc072 Initial revision
glantau
parents:
diff changeset
802 #define op_avg(a, b) a = avg2(a, b)
986e461dc072 Initial revision
glantau
parents:
diff changeset
803 #define op_sub(a, b) a -= b
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
804 #define op_put(a, b) a = b
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
805
986e461dc072 Initial revision
glantau
parents:
diff changeset
806 PIXOP(DCTELEM, sub, op_sub, 8)
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
807 PIXOP(uint8_t, avg, op_avg, line_size)
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
808 PIXOP(uint8_t, put, op_put, line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
809
986e461dc072 Initial revision
glantau
parents:
diff changeset
810 /* not rounding primitives */
986e461dc072 Initial revision
glantau
parents:
diff changeset
811 #undef avg2
986e461dc072 Initial revision
glantau
parents:
diff changeset
812 #undef avg4
986e461dc072 Initial revision
glantau
parents:
diff changeset
813 #define avg2(a,b) ((a+b)>>1)
986e461dc072 Initial revision
glantau
parents:
diff changeset
814 #define avg4(a,b,c,d) ((a+b+c+d+1)>>2)
986e461dc072 Initial revision
glantau
parents:
diff changeset
815
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
816 PIXOP(uint8_t, avg_no_rnd, op_avg, line_size)
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
817 PIXOP(uint8_t, put_no_rnd, op_put, line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
818 /* motion estimation */
986e461dc072 Initial revision
glantau
parents:
diff changeset
819
986e461dc072 Initial revision
glantau
parents:
diff changeset
820 #undef avg2
986e461dc072 Initial revision
glantau
parents:
diff changeset
821 #undef avg4
403
2c3e25f4c496 removed unused stuff - added dsputil_set_bit_exact() support for easier testing
glantau
parents: 398
diff changeset
822 #endif
2c3e25f4c496 removed unused stuff - added dsputil_set_bit_exact() support for easier testing
glantau
parents: 398
diff changeset
823
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
824 #define avg2(a,b) ((a+b+1)>>1)
986e461dc072 Initial revision
glantau
parents:
diff changeset
825 #define avg4(a,b,c,d) ((a+b+c+d+2)>>2)
986e461dc072 Initial revision
glantau
parents:
diff changeset
826
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
827
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
828 static void gmc1_c(UINT8 *dst, UINT8 *src, int stride, int h, int x16, int y16, int rounder)
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
829 {
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
830 const int A=(16-x16)*(16-y16);
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
831 const int B=( x16)*(16-y16);
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
832 const int C=(16-x16)*( y16);
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
833 const int D=( x16)*( y16);
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
834 int i;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
835
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
836 for(i=0; i<h; i++)
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
837 {
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
838 dst[0]= (A*src[0] + B*src[1] + C*src[stride+0] + D*src[stride+1] + rounder)>>8;
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
839 dst[1]= (A*src[1] + B*src[2] + C*src[stride+1] + D*src[stride+2] + rounder)>>8;
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
840 dst[2]= (A*src[2] + B*src[3] + C*src[stride+2] + D*src[stride+3] + rounder)>>8;
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
841 dst[3]= (A*src[3] + B*src[4] + C*src[stride+3] + D*src[stride+4] + rounder)>>8;
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
842 dst[4]= (A*src[4] + B*src[5] + C*src[stride+4] + D*src[stride+5] + rounder)>>8;
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
843 dst[5]= (A*src[5] + B*src[6] + C*src[stride+5] + D*src[stride+6] + rounder)>>8;
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
844 dst[6]= (A*src[6] + B*src[7] + C*src[stride+6] + D*src[stride+7] + rounder)>>8;
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
845 dst[7]= (A*src[7] + B*src[8] + C*src[stride+7] + D*src[stride+8] + rounder)>>8;
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
846 dst+= stride;
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
847 src+= stride;
255
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
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
851 static void gmc_c(UINT8 *dst, UINT8 *src, int stride, int h, int ox, int oy,
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
852 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height)
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
853 {
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
854 int y, vx, vy;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
855 const int s= 1<<shift;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
856
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
857 width--;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
858 height--;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
859
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
860 for(y=0; y<h; y++){
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
861 int x;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
862
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
863 vx= ox;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
864 vy= oy;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
865 for(x=0; x<8; x++){ //XXX FIXME optimize
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
866 int src_x, src_y, frac_x, frac_y, index;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
867
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
868 src_x= vx>>16;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
869 src_y= vy>>16;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
870 frac_x= src_x&(s-1);
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
871 frac_y= src_y&(s-1);
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
872 src_x>>=shift;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
873 src_y>>=shift;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
874
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
875 if((unsigned)src_x < width){
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
876 if((unsigned)src_y < height){
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
877 index= src_x + src_y*stride;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
878 dst[y*stride + x]= ( ( src[index ]*(s-frac_x)
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
879 + src[index +1]* frac_x )*(s-frac_y)
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
880 + ( src[index+stride ]*(s-frac_x)
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
881 + src[index+stride+1]* frac_x )* frac_y
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
882 + r)>>(shift*2);
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
883 }else{
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
884 index= src_x + clip(src_y, 0, height)*stride;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
885 dst[y*stride + x]= ( ( src[index ]*(s-frac_x)
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
886 + src[index +1]* frac_x )*s
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
887 + r)>>(shift*2);
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
888 }
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
889 }else{
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
890 if((unsigned)src_y < height){
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
891 index= clip(src_x, 0, width) + src_y*stride;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
892 dst[y*stride + x]= ( ( src[index ]*(s-frac_y)
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
893 + src[index+stride ]* frac_y )*s
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
894 + r)>>(shift*2);
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
895 }else{
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
896 index= clip(src_x, 0, width) + clip(src_y, 0, height)*stride;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
897 dst[y*stride + x]= src[index ];
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
898 }
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
899 }
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
900
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
901 vx+= dxx;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
902 vy+= dyx;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
903 }
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
904 ox += dxy;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
905 oy += dyy;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
906 }
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
907 }
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
908
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
909 static inline void copy_block17(UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int h)
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
910 {
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
911 int i;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
912 for(i=0; i<h; i++)
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
913 {
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
914 ST32(dst , LD32(src ));
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
915 ST32(dst+4 , LD32(src+4 ));
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
916 ST32(dst+8 , LD32(src+8 ));
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
917 ST32(dst+12, LD32(src+12));
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
918 dst[16]= src[16];
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
919 dst+=dstStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
920 src+=srcStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
921 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
922 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
923
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
924 static inline void copy_block9(UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int h)
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
925 {
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
926 int i;
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
927 for(i=0; i<h; i++)
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
928 {
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
929 ST32(dst , LD32(src ));
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
930 ST32(dst+4 , LD32(src+4 ));
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
931 dst[8]= src[8];
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
932 dst+=dstStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
933 src+=srcStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
934 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
935 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
936
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
937 #define QPEL_MC(r, OPNAME, RND, OP) \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
938 static void OPNAME ## mpeg4_qpel8_h_lowpass(UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int h){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
939 UINT8 *cm = cropTbl + MAX_NEG_CROP;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
940 int i;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
941 for(i=0; i<h; i++)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
942 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
943 OP(dst[0], (src[0]+src[1])*20 - (src[0]+src[2])*6 + (src[1]+src[3])*3 - (src[2]+src[4]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
944 OP(dst[1], (src[1]+src[2])*20 - (src[0]+src[3])*6 + (src[0]+src[4])*3 - (src[1]+src[5]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
945 OP(dst[2], (src[2]+src[3])*20 - (src[1]+src[4])*6 + (src[0]+src[5])*3 - (src[0]+src[6]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
946 OP(dst[3], (src[3]+src[4])*20 - (src[2]+src[5])*6 + (src[1]+src[6])*3 - (src[0]+src[7]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
947 OP(dst[4], (src[4]+src[5])*20 - (src[3]+src[6])*6 + (src[2]+src[7])*3 - (src[1]+src[8]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
948 OP(dst[5], (src[5]+src[6])*20 - (src[4]+src[7])*6 + (src[3]+src[8])*3 - (src[2]+src[8]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
949 OP(dst[6], (src[6]+src[7])*20 - (src[5]+src[8])*6 + (src[4]+src[8])*3 - (src[3]+src[7]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
950 OP(dst[7], (src[7]+src[8])*20 - (src[6]+src[8])*6 + (src[5]+src[7])*3 - (src[4]+src[6]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
951 dst+=dstStride;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
952 src+=srcStride;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
953 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
954 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
955 \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
956 static void OPNAME ## mpeg4_qpel8_v_lowpass(UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int w){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
957 UINT8 *cm = cropTbl + MAX_NEG_CROP;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
958 int i;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
959 for(i=0; i<w; i++)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
960 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
961 const int src0= src[0*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
962 const int src1= src[1*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
963 const int src2= src[2*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
964 const int src3= src[3*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
965 const int src4= src[4*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
966 const int src5= src[5*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
967 const int src6= src[6*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
968 const int src7= src[7*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
969 const int src8= src[8*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
970 OP(dst[0*dstStride], (src0+src1)*20 - (src0+src2)*6 + (src1+src3)*3 - (src2+src4));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
971 OP(dst[1*dstStride], (src1+src2)*20 - (src0+src3)*6 + (src0+src4)*3 - (src1+src5));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
972 OP(dst[2*dstStride], (src2+src3)*20 - (src1+src4)*6 + (src0+src5)*3 - (src0+src6));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
973 OP(dst[3*dstStride], (src3+src4)*20 - (src2+src5)*6 + (src1+src6)*3 - (src0+src7));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
974 OP(dst[4*dstStride], (src4+src5)*20 - (src3+src6)*6 + (src2+src7)*3 - (src1+src8));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
975 OP(dst[5*dstStride], (src5+src6)*20 - (src4+src7)*6 + (src3+src8)*3 - (src2+src8));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
976 OP(dst[6*dstStride], (src6+src7)*20 - (src5+src8)*6 + (src4+src8)*3 - (src3+src7));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
977 OP(dst[7*dstStride], (src7+src8)*20 - (src6+src8)*6 + (src5+src7)*3 - (src4+src6));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
978 dst++;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
979 src++;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
980 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
981 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
982 \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
983 static void OPNAME ## mpeg4_qpel16_h_lowpass(UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int h){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
984 UINT8 *cm = cropTbl + MAX_NEG_CROP;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
985 int i;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
986 for(i=0; i<h; i++)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
987 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
988 OP(dst[ 0], (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
989 OP(dst[ 1], (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
990 OP(dst[ 2], (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
991 OP(dst[ 3], (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
992 OP(dst[ 4], (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
993 OP(dst[ 5], (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 9]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
994 OP(dst[ 6], (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 9])*3 - (src[ 3]+src[10]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
995 OP(dst[ 7], (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 9])*6 + (src[ 5]+src[10])*3 - (src[ 4]+src[11]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
996 OP(dst[ 8], (src[ 8]+src[ 9])*20 - (src[ 7]+src[10])*6 + (src[ 6]+src[11])*3 - (src[ 5]+src[12]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
997 OP(dst[ 9], (src[ 9]+src[10])*20 - (src[ 8]+src[11])*6 + (src[ 7]+src[12])*3 - (src[ 6]+src[13]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
998 OP(dst[10], (src[10]+src[11])*20 - (src[ 9]+src[12])*6 + (src[ 8]+src[13])*3 - (src[ 7]+src[14]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
999 OP(dst[11], (src[11]+src[12])*20 - (src[10]+src[13])*6 + (src[ 9]+src[14])*3 - (src[ 8]+src[15]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1000 OP(dst[12], (src[12]+src[13])*20 - (src[11]+src[14])*6 + (src[10]+src[15])*3 - (src[ 9]+src[16]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1001 OP(dst[13], (src[13]+src[14])*20 - (src[12]+src[15])*6 + (src[11]+src[16])*3 - (src[10]+src[16]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1002 OP(dst[14], (src[14]+src[15])*20 - (src[13]+src[16])*6 + (src[12]+src[16])*3 - (src[11]+src[15]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1003 OP(dst[15], (src[15]+src[16])*20 - (src[14]+src[16])*6 + (src[13]+src[15])*3 - (src[12]+src[14]));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1004 dst+=dstStride;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1005 src+=srcStride;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1006 }\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1007 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1008 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1009 static void OPNAME ## mpeg4_qpel16_v_lowpass(UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int w){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1010 UINT8 *cm = cropTbl + MAX_NEG_CROP;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1011 int i;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1012 for(i=0; i<w; i++)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1013 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1014 const int src0= src[0*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1015 const int src1= src[1*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1016 const int src2= src[2*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1017 const int src3= src[3*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1018 const int src4= src[4*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1019 const int src5= src[5*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1020 const int src6= src[6*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1021 const int src7= src[7*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1022 const int src8= src[8*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1023 const int src9= src[9*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1024 const int src10= src[10*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1025 const int src11= src[11*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1026 const int src12= src[12*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1027 const int src13= src[13*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1028 const int src14= src[14*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1029 const int src15= src[15*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1030 const int src16= src[16*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1031 OP(dst[ 0*dstStride], (src0 +src1 )*20 - (src0 +src2 )*6 + (src1 +src3 )*3 - (src2 +src4 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1032 OP(dst[ 1*dstStride], (src1 +src2 )*20 - (src0 +src3 )*6 + (src0 +src4 )*3 - (src1 +src5 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1033 OP(dst[ 2*dstStride], (src2 +src3 )*20 - (src1 +src4 )*6 + (src0 +src5 )*3 - (src0 +src6 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1034 OP(dst[ 3*dstStride], (src3 +src4 )*20 - (src2 +src5 )*6 + (src1 +src6 )*3 - (src0 +src7 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1035 OP(dst[ 4*dstStride], (src4 +src5 )*20 - (src3 +src6 )*6 + (src2 +src7 )*3 - (src1 +src8 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1036 OP(dst[ 5*dstStride], (src5 +src6 )*20 - (src4 +src7 )*6 + (src3 +src8 )*3 - (src2 +src9 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1037 OP(dst[ 6*dstStride], (src6 +src7 )*20 - (src5 +src8 )*6 + (src4 +src9 )*3 - (src3 +src10));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1038 OP(dst[ 7*dstStride], (src7 +src8 )*20 - (src6 +src9 )*6 + (src5 +src10)*3 - (src4 +src11));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1039 OP(dst[ 8*dstStride], (src8 +src9 )*20 - (src7 +src10)*6 + (src6 +src11)*3 - (src5 +src12));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1040 OP(dst[ 9*dstStride], (src9 +src10)*20 - (src8 +src11)*6 + (src7 +src12)*3 - (src6 +src13));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1041 OP(dst[10*dstStride], (src10+src11)*20 - (src9 +src12)*6 + (src8 +src13)*3 - (src7 +src14));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1042 OP(dst[11*dstStride], (src11+src12)*20 - (src10+src13)*6 + (src9 +src14)*3 - (src8 +src15));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1043 OP(dst[12*dstStride], (src12+src13)*20 - (src11+src14)*6 + (src10+src15)*3 - (src9 +src16));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1044 OP(dst[13*dstStride], (src13+src14)*20 - (src12+src15)*6 + (src11+src16)*3 - (src10+src16));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1045 OP(dst[14*dstStride], (src14+src15)*20 - (src13+src16)*6 + (src12+src16)*3 - (src11+src15));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1046 OP(dst[15*dstStride], (src15+src16)*20 - (src14+src16)*6 + (src13+src15)*3 - (src12+src14));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1047 dst++;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1048 src++;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1049 }\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1050 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1051 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1052 static void OPNAME ## qpel8_mc00_c (UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1053 OPNAME ## pixels8(dst, src, stride, 8);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1054 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1055 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1056 static void OPNAME ## qpel8_mc10_c(UINT8 *dst, UINT8 *src, int stride){\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1057 UINT8 half[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1058 put ## RND ## mpeg4_qpel8_h_lowpass(half, src, 8, stride, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1059 OPNAME ## pixels8_l2(dst, src, half, stride, stride, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1060 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1061 \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1062 static void OPNAME ## qpel8_mc20_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1063 OPNAME ## mpeg4_qpel8_h_lowpass(dst, src, stride, stride, 8);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1064 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1065 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1066 static void OPNAME ## qpel8_mc30_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1067 UINT8 half[64];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1068 put ## RND ## mpeg4_qpel8_h_lowpass(half, src, 8, stride, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1069 OPNAME ## pixels8_l2(dst, src+1, half, stride, stride, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1070 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1071 \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1072 static void OPNAME ## qpel8_mc01_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1073 UINT8 full[16*9];\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1074 UINT8 half[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1075 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1076 put ## RND ## mpeg4_qpel8_v_lowpass(half, full, 8, 16, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1077 OPNAME ## pixels8_l2(dst, full, half, stride, 16, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1078 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1079 \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1080 static void OPNAME ## qpel8_mc02_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1081 UINT8 full[16*9];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1082 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1083 OPNAME ## mpeg4_qpel8_v_lowpass(dst, full, stride, 16, 8);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1084 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1085 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1086 static void OPNAME ## qpel8_mc03_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1087 UINT8 full[16*9];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1088 UINT8 half[64];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1089 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1090 put ## RND ## mpeg4_qpel8_v_lowpass(half, full, 8, 16, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1091 OPNAME ## pixels8_l2(dst, full+16, half, stride, 16, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1092 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1093 static void OPNAME ## qpel8_mc11_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1094 UINT8 full[16*9];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1095 UINT8 halfH[72];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1096 UINT8 halfV[64];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1097 UINT8 halfHV[64];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1098 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1099 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1100 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full, 8, 16, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1101 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1102 OPNAME ## pixels8_l4(dst, full, halfH, halfV, halfHV, stride, 16, 8, 8, 8, 8);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1103 }\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1104 static void OPNAME ## qpel8_mc31_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1105 UINT8 full[16*9];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1106 UINT8 halfH[72];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1107 UINT8 halfV[64];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1108 UINT8 halfHV[64];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1109 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1110 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1111 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full+1, 8, 16, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1112 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1113 OPNAME ## pixels8_l4(dst, full+1, halfH, halfV, halfHV, stride, 16, 8, 8, 8, 8);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1114 }\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1115 static void OPNAME ## qpel8_mc13_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1116 UINT8 full[16*9];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1117 UINT8 halfH[72];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1118 UINT8 halfV[64];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1119 UINT8 halfHV[64];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1120 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1121 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1122 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full, 8, 16, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1123 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1124 OPNAME ## pixels8_l4(dst, full+16, halfH+8, halfV, halfHV, stride, 16, 8, 8, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1125 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1126 static void OPNAME ## qpel8_mc33_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1127 UINT8 full[16*9];\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1128 UINT8 halfH[72];\
256
4c1cec7c3c7c q-pel mc fixed
michaelni
parents: 255
diff changeset
1129 UINT8 halfV[64];\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1130 UINT8 halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1131 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1132 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full , 8, 16, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1133 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full+1, 8, 16, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1134 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1135 OPNAME ## pixels8_l4(dst, full+17, halfH+8, halfV, halfHV, stride, 16, 8, 8, 8, 8);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1136 }\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1137 static void OPNAME ## qpel8_mc21_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1138 UINT8 halfH[72];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1139 UINT8 halfHV[64];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1140 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, src, 8, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1141 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1142 OPNAME ## pixels8_l2(dst, halfH, halfHV, stride, 8, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1143 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1144 static void OPNAME ## qpel8_mc23_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1145 UINT8 halfH[72];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1146 UINT8 halfHV[64];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1147 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, src, 8, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1148 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1149 OPNAME ## pixels8_l2(dst, halfH+8, halfHV, stride, 8, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1150 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1151 static void OPNAME ## qpel8_mc12_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1152 UINT8 full[16*9];\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1153 UINT8 halfH[72];\
256
4c1cec7c3c7c q-pel mc fixed
michaelni
parents: 255
diff changeset
1154 UINT8 halfV[64];\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1155 UINT8 halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1156 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1157 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1158 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full, 8, 16, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1159 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1160 OPNAME ## pixels8_l2(dst, halfV, halfHV, stride, 8, 8, 8);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1161 }\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1162 static void OPNAME ## qpel8_mc32_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1163 UINT8 full[16*9];\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1164 UINT8 halfH[72];\
256
4c1cec7c3c7c q-pel mc fixed
michaelni
parents: 255
diff changeset
1165 UINT8 halfV[64];\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1166 UINT8 halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1167 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1168 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1169 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full+1, 8, 16, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1170 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1171 OPNAME ## pixels8_l2(dst, halfV, halfHV, stride, 8, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1172 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1173 static void OPNAME ## qpel8_mc22_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1174 UINT8 halfH[72];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1175 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, src, 8, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1176 OPNAME ## mpeg4_qpel8_v_lowpass(dst, halfH, stride, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1177 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1178 static void OPNAME ## qpel16_mc00_c (UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1179 OPNAME ## pixels16(dst, src, stride, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1180 }\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1181 \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1182 static void OPNAME ## qpel16_mc10_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1183 UINT8 half[256];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1184 put ## RND ## mpeg4_qpel16_h_lowpass(half, src, 16, stride, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1185 OPNAME ## pixels16_l2(dst, src, half, stride, stride, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1186 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1187 \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1188 static void OPNAME ## qpel16_mc20_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1189 OPNAME ## mpeg4_qpel16_h_lowpass(dst, src, stride, stride, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1190 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1191 \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1192 static void OPNAME ## qpel16_mc30_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1193 UINT8 half[256];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1194 put ## RND ## mpeg4_qpel16_h_lowpass(half, src, 16, stride, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1195 OPNAME ## pixels16_l2(dst, src+1, half, stride, stride, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1196 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1197 \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1198 static void OPNAME ## qpel16_mc01_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1199 UINT8 full[24*17];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1200 UINT8 half[256];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1201 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1202 put ## RND ## mpeg4_qpel16_v_lowpass(half, full, 16, 24, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1203 OPNAME ## pixels16_l2(dst, full, half, stride, 24, 16, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1204 }\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1205 \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1206 static void OPNAME ## qpel16_mc02_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1207 UINT8 full[24*17];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1208 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1209 OPNAME ## mpeg4_qpel16_v_lowpass(dst, full, stride, 24, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1210 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1211 \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1212 static void OPNAME ## qpel16_mc03_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1213 UINT8 full[24*17];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1214 UINT8 half[256];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1215 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1216 put ## RND ## mpeg4_qpel16_v_lowpass(half, full, 16, 24, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1217 OPNAME ## pixels16_l2(dst, full+24, half, stride, 24, 16, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1218 }\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1219 static void OPNAME ## qpel16_mc11_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1220 UINT8 full[24*17];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1221 UINT8 halfH[272];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1222 UINT8 halfV[256];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1223 UINT8 halfHV[256];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1224 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1225 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1226 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full, 16, 24, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1227 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1228 OPNAME ## pixels16_l4(dst, full, halfH, halfV, halfHV, stride, 24, 16, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1229 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1230 static void OPNAME ## qpel16_mc31_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1231 UINT8 full[24*17];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1232 UINT8 halfH[272];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1233 UINT8 halfV[256];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1234 UINT8 halfHV[256];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1235 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1236 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1237 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full+1, 16, 24, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1238 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1239 OPNAME ## pixels16_l4(dst, full+1, halfH, halfV, halfHV, stride, 24, 16, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1240 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1241 static void OPNAME ## qpel16_mc13_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1242 UINT8 full[24*17];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1243 UINT8 halfH[272];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1244 UINT8 halfV[256];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1245 UINT8 halfHV[256];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1246 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1247 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1248 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full, 16, 24, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1249 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1250 OPNAME ## pixels16_l4(dst, full+24, halfH+16, halfV, halfHV, stride, 24, 16, 16, 16, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1251 }\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1252 static void OPNAME ## qpel16_mc33_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1253 UINT8 full[24*17];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1254 UINT8 halfH[272];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1255 UINT8 halfV[256];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1256 UINT8 halfHV[256];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1257 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1258 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full , 16, 24, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1259 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full+1, 16, 24, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1260 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1261 OPNAME ## pixels16_l4(dst, full+25, halfH+16, halfV, halfHV, stride, 24, 16, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1262 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1263 static void OPNAME ## qpel16_mc21_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1264 UINT8 halfH[272];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1265 UINT8 halfHV[256];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1266 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, src, 16, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1267 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1268 OPNAME ## pixels16_l2(dst, halfH, halfHV, stride, 16, 16, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1269 }\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1270 static void OPNAME ## qpel16_mc23_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1271 UINT8 halfH[272];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1272 UINT8 halfHV[256];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1273 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, src, 16, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1274 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1275 OPNAME ## pixels16_l2(dst, halfH+16, halfHV, stride, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1276 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1277 static void OPNAME ## qpel16_mc12_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1278 UINT8 full[24*17];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1279 UINT8 halfH[272];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1280 UINT8 halfV[256];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1281 UINT8 halfHV[256];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1282 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1283 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1284 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full, 16, 24, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1285 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1286 OPNAME ## pixels16_l2(dst, halfV, halfHV, stride, 16, 16, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1287 }\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1288 static void OPNAME ## qpel16_mc32_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1289 UINT8 full[24*17];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1290 UINT8 halfH[272];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1291 UINT8 halfV[256];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1292 UINT8 halfHV[256];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1293 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1294 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1295 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full+1, 16, 24, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1296 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1297 OPNAME ## pixels16_l2(dst, halfV, halfHV, stride, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1298 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1299 static void OPNAME ## qpel16_mc22_c(UINT8 *dst, UINT8 *src, int stride){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1300 UINT8 halfH[272];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1301 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, src, 16, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1302 OPNAME ## mpeg4_qpel16_v_lowpass(dst, halfH, stride, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1303 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1304 qpel_mc_func OPNAME ## qpel_pixels_tab[2][16]={ \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1305 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1306 OPNAME ## qpel16_mc00_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1307 OPNAME ## qpel16_mc10_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1308 OPNAME ## qpel16_mc20_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1309 OPNAME ## qpel16_mc30_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1310 OPNAME ## qpel16_mc01_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1311 OPNAME ## qpel16_mc11_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1312 OPNAME ## qpel16_mc21_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1313 OPNAME ## qpel16_mc31_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1314 OPNAME ## qpel16_mc02_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1315 OPNAME ## qpel16_mc12_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1316 OPNAME ## qpel16_mc22_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1317 OPNAME ## qpel16_mc32_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1318 OPNAME ## qpel16_mc03_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1319 OPNAME ## qpel16_mc13_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1320 OPNAME ## qpel16_mc23_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1321 OPNAME ## qpel16_mc33_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1322 },{\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1323 OPNAME ## qpel8_mc00_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1324 OPNAME ## qpel8_mc10_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1325 OPNAME ## qpel8_mc20_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1326 OPNAME ## qpel8_mc30_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1327 OPNAME ## qpel8_mc01_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1328 OPNAME ## qpel8_mc11_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1329 OPNAME ## qpel8_mc21_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1330 OPNAME ## qpel8_mc31_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1331 OPNAME ## qpel8_mc02_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1332 OPNAME ## qpel8_mc12_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1333 OPNAME ## qpel8_mc22_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1334 OPNAME ## qpel8_mc32_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1335 OPNAME ## qpel8_mc03_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1336 OPNAME ## qpel8_mc13_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1337 OPNAME ## qpel8_mc23_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1338 OPNAME ## qpel8_mc33_c, \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1339 }\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1340 };
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1341
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1342 #define op_avg(a, b) a = (((a)+cm[((b) + 16)>>5]+1)>>1)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1343 #define op_avg_no_rnd(a, b) a = (((a)+cm[((b) + 15)>>5])>>1)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1344 #define op_put(a, b) a = cm[((b) + 16)>>5]
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1345 #define op_put_no_rnd(a, b) a = cm[((b) + 15)>>5]
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1346
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1347 QPEL_MC(0, put_ , _ , op_put)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1348 QPEL_MC(1, put_no_rnd_, _no_rnd_, op_put_no_rnd)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1349 QPEL_MC(0, avg_ , _ , op_avg)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1350 //QPEL_MC(1, avg_no_rnd , _ , op_avg)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1351 #undef op_avg
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1352 #undef op_avg_no_rnd
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1353 #undef op_put
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1354 #undef op_put_no_rnd
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1355
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1356 int pix_abs16x16_c(UINT8 *pix1, UINT8 *pix2, int line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1357 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
1358 int s, i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1359
986e461dc072 Initial revision
glantau
parents:
diff changeset
1360 s = 0;
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1361 for(i=0;i<16;i++) {
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1362 s += abs(pix1[0] - pix2[0]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1363 s += abs(pix1[1] - pix2[1]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1364 s += abs(pix1[2] - pix2[2]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1365 s += abs(pix1[3] - pix2[3]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1366 s += abs(pix1[4] - pix2[4]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1367 s += abs(pix1[5] - pix2[5]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1368 s += abs(pix1[6] - pix2[6]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1369 s += abs(pix1[7] - pix2[7]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1370 s += abs(pix1[8] - pix2[8]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1371 s += abs(pix1[9] - pix2[9]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1372 s += abs(pix1[10] - pix2[10]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1373 s += abs(pix1[11] - pix2[11]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1374 s += abs(pix1[12] - pix2[12]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1375 s += abs(pix1[13] - pix2[13]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1376 s += abs(pix1[14] - pix2[14]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1377 s += abs(pix1[15] - pix2[15]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1378 pix1 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1379 pix2 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1380 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
1381 return s;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1382 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
1383
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1384 int pix_abs16x16_x2_c(UINT8 *pix1, UINT8 *pix2, int line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1385 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
1386 int s, i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1387
986e461dc072 Initial revision
glantau
parents:
diff changeset
1388 s = 0;
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1389 for(i=0;i<16;i++) {
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1390 s += abs(pix1[0] - avg2(pix2[0], pix2[1]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1391 s += abs(pix1[1] - avg2(pix2[1], pix2[2]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1392 s += abs(pix1[2] - avg2(pix2[2], pix2[3]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1393 s += abs(pix1[3] - avg2(pix2[3], pix2[4]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1394 s += abs(pix1[4] - avg2(pix2[4], pix2[5]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1395 s += abs(pix1[5] - avg2(pix2[5], pix2[6]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1396 s += abs(pix1[6] - avg2(pix2[6], pix2[7]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1397 s += abs(pix1[7] - avg2(pix2[7], pix2[8]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1398 s += abs(pix1[8] - avg2(pix2[8], pix2[9]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1399 s += abs(pix1[9] - avg2(pix2[9], pix2[10]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1400 s += abs(pix1[10] - avg2(pix2[10], pix2[11]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1401 s += abs(pix1[11] - avg2(pix2[11], pix2[12]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1402 s += abs(pix1[12] - avg2(pix2[12], pix2[13]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1403 s += abs(pix1[13] - avg2(pix2[13], pix2[14]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1404 s += abs(pix1[14] - avg2(pix2[14], pix2[15]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1405 s += abs(pix1[15] - avg2(pix2[15], pix2[16]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1406 pix1 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1407 pix2 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1408 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
1409 return s;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1410 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
1411
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1412 int pix_abs16x16_y2_c(UINT8 *pix1, UINT8 *pix2, int line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1413 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
1414 int s, i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1415 UINT8 *pix3 = pix2 + line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1416
986e461dc072 Initial revision
glantau
parents:
diff changeset
1417 s = 0;
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1418 for(i=0;i<16;i++) {
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1419 s += abs(pix1[0] - avg2(pix2[0], pix3[0]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1420 s += abs(pix1[1] - avg2(pix2[1], pix3[1]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1421 s += abs(pix1[2] - avg2(pix2[2], pix3[2]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1422 s += abs(pix1[3] - avg2(pix2[3], pix3[3]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1423 s += abs(pix1[4] - avg2(pix2[4], pix3[4]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1424 s += abs(pix1[5] - avg2(pix2[5], pix3[5]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1425 s += abs(pix1[6] - avg2(pix2[6], pix3[6]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1426 s += abs(pix1[7] - avg2(pix2[7], pix3[7]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1427 s += abs(pix1[8] - avg2(pix2[8], pix3[8]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1428 s += abs(pix1[9] - avg2(pix2[9], pix3[9]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1429 s += abs(pix1[10] - avg2(pix2[10], pix3[10]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1430 s += abs(pix1[11] - avg2(pix2[11], pix3[11]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1431 s += abs(pix1[12] - avg2(pix2[12], pix3[12]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1432 s += abs(pix1[13] - avg2(pix2[13], pix3[13]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1433 s += abs(pix1[14] - avg2(pix2[14], pix3[14]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1434 s += abs(pix1[15] - avg2(pix2[15], pix3[15]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1435 pix1 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1436 pix2 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1437 pix3 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1438 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
1439 return s;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1440 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
1441
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1442 int pix_abs16x16_xy2_c(UINT8 *pix1, UINT8 *pix2, int line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1443 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
1444 int s, i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1445 UINT8 *pix3 = pix2 + line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1446
986e461dc072 Initial revision
glantau
parents:
diff changeset
1447 s = 0;
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1448 for(i=0;i<16;i++) {
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1449 s += abs(pix1[0] - avg4(pix2[0], pix2[1], pix3[0], pix3[1]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1450 s += abs(pix1[1] - avg4(pix2[1], pix2[2], pix3[1], pix3[2]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1451 s += abs(pix1[2] - avg4(pix2[2], pix2[3], pix3[2], pix3[3]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1452 s += abs(pix1[3] - avg4(pix2[3], pix2[4], pix3[3], pix3[4]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1453 s += abs(pix1[4] - avg4(pix2[4], pix2[5], pix3[4], pix3[5]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1454 s += abs(pix1[5] - avg4(pix2[5], pix2[6], pix3[5], pix3[6]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1455 s += abs(pix1[6] - avg4(pix2[6], pix2[7], pix3[6], pix3[7]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1456 s += abs(pix1[7] - avg4(pix2[7], pix2[8], pix3[7], pix3[8]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1457 s += abs(pix1[8] - avg4(pix2[8], pix2[9], pix3[8], pix3[9]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1458 s += abs(pix1[9] - avg4(pix2[9], pix2[10], pix3[9], pix3[10]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1459 s += abs(pix1[10] - avg4(pix2[10], pix2[11], pix3[10], pix3[11]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1460 s += abs(pix1[11] - avg4(pix2[11], pix2[12], pix3[11], pix3[12]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1461 s += abs(pix1[12] - avg4(pix2[12], pix2[13], pix3[12], pix3[13]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1462 s += abs(pix1[13] - avg4(pix2[13], pix2[14], pix3[13], pix3[14]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1463 s += abs(pix1[14] - avg4(pix2[14], pix2[15], pix3[14], pix3[15]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1464 s += abs(pix1[15] - avg4(pix2[15], pix2[16], pix3[15], pix3[16]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
1465 pix1 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1466 pix2 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1467 pix3 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1468 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
1469 return s;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1470 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
1471
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1472 int pix_abs8x8_c(UINT8 *pix1, UINT8 *pix2, int line_size)
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1473 {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1474 int s, i;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1475
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1476 s = 0;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1477 for(i=0;i<8;i++) {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1478 s += abs(pix1[0] - pix2[0]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1479 s += abs(pix1[1] - pix2[1]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1480 s += abs(pix1[2] - pix2[2]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1481 s += abs(pix1[3] - pix2[3]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1482 s += abs(pix1[4] - pix2[4]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1483 s += abs(pix1[5] - pix2[5]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1484 s += abs(pix1[6] - pix2[6]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1485 s += abs(pix1[7] - pix2[7]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1486 pix1 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1487 pix2 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1488 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1489 return s;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1490 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1491
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1492 int pix_abs8x8_x2_c(UINT8 *pix1, UINT8 *pix2, int line_size)
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1493 {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1494 int s, i;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1495
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1496 s = 0;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1497 for(i=0;i<8;i++) {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1498 s += abs(pix1[0] - avg2(pix2[0], pix2[1]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1499 s += abs(pix1[1] - avg2(pix2[1], pix2[2]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1500 s += abs(pix1[2] - avg2(pix2[2], pix2[3]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1501 s += abs(pix1[3] - avg2(pix2[3], pix2[4]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1502 s += abs(pix1[4] - avg2(pix2[4], pix2[5]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1503 s += abs(pix1[5] - avg2(pix2[5], pix2[6]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1504 s += abs(pix1[6] - avg2(pix2[6], pix2[7]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1505 s += abs(pix1[7] - avg2(pix2[7], pix2[8]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1506 pix1 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1507 pix2 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1508 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1509 return s;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1510 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1511
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1512 int pix_abs8x8_y2_c(UINT8 *pix1, UINT8 *pix2, int line_size)
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1513 {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1514 int s, i;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1515 UINT8 *pix3 = pix2 + line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1516
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1517 s = 0;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1518 for(i=0;i<8;i++) {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1519 s += abs(pix1[0] - avg2(pix2[0], pix3[0]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1520 s += abs(pix1[1] - avg2(pix2[1], pix3[1]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1521 s += abs(pix1[2] - avg2(pix2[2], pix3[2]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1522 s += abs(pix1[3] - avg2(pix2[3], pix3[3]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1523 s += abs(pix1[4] - avg2(pix2[4], pix3[4]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1524 s += abs(pix1[5] - avg2(pix2[5], pix3[5]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1525 s += abs(pix1[6] - avg2(pix2[6], pix3[6]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1526 s += abs(pix1[7] - avg2(pix2[7], pix3[7]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1527 pix1 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1528 pix2 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1529 pix3 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1530 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1531 return s;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1532 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1533
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1534 int pix_abs8x8_xy2_c(UINT8 *pix1, UINT8 *pix2, int line_size)
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1535 {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1536 int s, i;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1537 UINT8 *pix3 = pix2 + line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1538
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1539 s = 0;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1540 for(i=0;i<8;i++) {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1541 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
1542 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
1543 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
1544 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
1545 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
1546 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
1547 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
1548 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
1549 pix1 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1550 pix2 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1551 pix3 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1552 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1553 return s;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1554 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1555
764
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
1556 void ff_block_permute(INT16 *block, UINT8 *permutation, const UINT8 *scantable, int last)
174
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1557 {
764
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
1558 int i;
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
1559 INT16 temp[64];
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
1560
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
1561 if(last<=0) return;
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
1562 if(permutation[1]==1) return; //FIXME its ok but not clean and might fail for some perms
174
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1563
764
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
1564 for(i=0; i<=last; i++){
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
1565 const int j= scantable[i];
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
1566 temp[j]= block[j];
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
1567 block[j]=0;
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
1568 }
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
1569
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
1570 for(i=0; i<=last; i++){
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
1571 const int j= scantable[i];
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
1572 const int perm_j= permutation[j];
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
1573 block[perm_j]= temp[j];
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
1574 }
174
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
1575 }
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
1576
296
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
1577 void clear_blocks_c(DCTELEM *blocks)
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
1578 {
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
1579 memset(blocks, 0, sizeof(DCTELEM)*6*64);
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
1580 }
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
1581
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1582 void dsputil_init(void)
986e461dc072 Initial revision
glantau
parents:
diff changeset
1583 {
751
cbe316f082bc warning fixes
michaelni
parents: 706
diff changeset
1584 int i;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1585
986e461dc072 Initial revision
glantau
parents:
diff changeset
1586 for(i=0;i<256;i++) cropTbl[i + MAX_NEG_CROP] = i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1587 for(i=0;i<MAX_NEG_CROP;i++) {
986e461dc072 Initial revision
glantau
parents:
diff changeset
1588 cropTbl[i] = 0;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1589 cropTbl[i + MAX_NEG_CROP + 256] = 255;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1590 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
1591
986e461dc072 Initial revision
glantau
parents:
diff changeset
1592 for(i=0;i<512;i++) {
986e461dc072 Initial revision
glantau
parents:
diff changeset
1593 squareTbl[i] = (i - 256) * (i - 256);
986e461dc072 Initial revision
glantau
parents:
diff changeset
1594 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
1595
986e461dc072 Initial revision
glantau
parents:
diff changeset
1596 get_pixels = get_pixels_c;
324
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
1597 diff_pixels = diff_pixels_c;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1598 put_pixels_clamped = put_pixels_clamped_c;
986e461dc072 Initial revision
glantau
parents:
diff changeset
1599 add_pixels_clamped = add_pixels_clamped_c;
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1600 ff_gmc1= gmc1_c;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1601 ff_gmc= gmc_c;
296
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
1602 clear_blocks= clear_blocks_c;
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
1603 pix_sum= pix_sum_c;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
1604 pix_norm1= pix_norm1_c;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1605
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1606 pix_abs16x16 = pix_abs16x16_c;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1607 pix_abs16x16_x2 = pix_abs16x16_x2_c;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1608 pix_abs16x16_y2 = pix_abs16x16_y2_c;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1609 pix_abs16x16_xy2 = pix_abs16x16_xy2_c;
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1610 pix_abs8x8 = pix_abs8x8_c;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1611 pix_abs8x8_x2 = pix_abs8x8_x2_c;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1612 pix_abs8x8_y2 = pix_abs8x8_y2_c;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
1613 pix_abs8x8_xy2 = pix_abs8x8_xy2_c;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1614
2
2e2c46c87460 fixed config for direct mplayer build compatibility
glantau
parents: 0
diff changeset
1615 #ifdef HAVE_MMX
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1616 dsputil_init_mmx();
986e461dc072 Initial revision
glantau
parents:
diff changeset
1617 #endif
62
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
1618 #ifdef ARCH_ARMV4L
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
1619 dsputil_init_armv4l();
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
1620 #endif
88
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1621 #ifdef HAVE_MLIB
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1622 dsputil_init_mlib();
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1623 #endif
214
73df666cacc7 Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents: 209
diff changeset
1624 #ifdef ARCH_ALPHA
73df666cacc7 Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents: 209
diff changeset
1625 dsputil_init_alpha();
73df666cacc7 Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents: 209
diff changeset
1626 #endif
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 612
diff changeset
1627 #ifdef ARCH_POWERPC
638
0012f75c92bb altivec build tidyup patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 626
diff changeset
1628 dsputil_init_ppc();
626
23a093d6e450 patch by Heliodoro Tammaro <helio at interactives dot org>
michaelni
parents: 625
diff changeset
1629 #endif
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 676
diff changeset
1630 #ifdef HAVE_MMI
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 676
diff changeset
1631 dsputil_init_mmi();
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 676
diff changeset
1632 #endif
88
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
1633
706
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 689
diff changeset
1634 for(i=0; i<64; i++) inv_zigzag_direct16[ff_zigzag_direct[i]]= i+1;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1635 }
252
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1636
403
2c3e25f4c496 removed unused stuff - added dsputil_set_bit_exact() support for easier testing
glantau
parents: 398
diff changeset
1637 /* 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
1638 void avcodec_set_bit_exact(void)
2c3e25f4c496 removed unused stuff - added dsputil_set_bit_exact() support for easier testing
glantau
parents: 398
diff changeset
1639 {
676
c3bdb00a98a9 dont store version for bit-exact tests
michaelni
parents: 651
diff changeset
1640 ff_bit_exact=1;
403
2c3e25f4c496 removed unused stuff - added dsputil_set_bit_exact() support for easier testing
glantau
parents: 398
diff changeset
1641 #ifdef HAVE_MMX
2c3e25f4c496 removed unused stuff - added dsputil_set_bit_exact() support for easier testing
glantau
parents: 398
diff changeset
1642 dsputil_set_bit_exact_mmx();
2c3e25f4c496 removed unused stuff - added dsputil_set_bit_exact() support for easier testing
glantau
parents: 398
diff changeset
1643 #endif
2c3e25f4c496 removed unused stuff - added dsputil_set_bit_exact() support for easier testing
glantau
parents: 398
diff changeset
1644 }
2c3e25f4c496 removed unused stuff - added dsputil_set_bit_exact() support for easier testing
glantau
parents: 398
diff changeset
1645
252
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1646 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
1647 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
1648 AVCodecContext *avctx)
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1649 {
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1650 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
1651 UINT8 *orig, *coded;
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1652 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
1653
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1654 quad = 0;
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1655 diff = 0;
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1656
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1657 /* Luminance */
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1658 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
1659 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
1660
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1661 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
1662 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
1663 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
1664 quad += sq[diff];
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1665 }
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1666 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
1667 coded += coded_linesize;
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1668 }
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1669
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1670 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
1671
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1672 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
1673 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
1674 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
1675 } else
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1676 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
1677 }
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
1678