annotate dsputil.c @ 2347:c6280d48be02 libavcodec

When bswap_32 is a macro, png images fail to decode properly, patch by (Milan Cutka <cutka>at<szm>dot<sk>)
author michael
date Fri, 12 Nov 2004 09:48:06 +0000
parents cd43603c46f9
children 23e7af611c34
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.
1739
07a484280a82 copyright year update of the files i touched and remembered, things look annoyingly unmaintained otherwise
michael
parents: 1729
diff changeset
4 * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
5 *
429
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
6 * This library is free software; you can redistribute it and/or
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
7 * modify it under the terms of the GNU Lesser General Public
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
8 * License as published by the Free Software Foundation; either
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
9 * version 2 of the License, or (at your option) any later version.
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
10 *
429
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
11 * This library is distributed in the hope that it will be useful,
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
429
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
14 * Lesser General Public License for more details.
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
15 *
429
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
16 * You should have received a copy of the GNU Lesser General Public
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
17 * License along with this library; if not, write to the Free Software
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
256
4c1cec7c3c7c q-pel mc fixed
michaelni
parents: 255
diff changeset
19 *
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
20 * gmc & q-pel & 32/64 bit based MC by Michael Niedermayer <michaelni@gmx.at>
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
21 */
1106
1e39f273ecd6 per file doxy
michaelni
parents: 1101
diff changeset
22
1e39f273ecd6 per file doxy
michaelni
parents: 1101
diff changeset
23 /**
1e39f273ecd6 per file doxy
michaelni
parents: 1101
diff changeset
24 * @file dsputil.c
1e39f273ecd6 per file doxy
michaelni
parents: 1101
diff changeset
25 * DSP utils
1e39f273ecd6 per file doxy
michaelni
parents: 1101
diff changeset
26 */
1e39f273ecd6 per file doxy
michaelni
parents: 1101
diff changeset
27
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
28 #include "avcodec.h"
986e461dc072 Initial revision
glantau
parents:
diff changeset
29 #include "dsputil.h"
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
30 #include "mpegvideo.h"
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
31 #include "simple_idct.h"
1557
5d53c03186a1 floating point AAN DCT
michael
parents: 1527
diff changeset
32 #include "faandct.h"
676
c3bdb00a98a9 dont store version for bit-exact tests
michaelni
parents: 651
diff changeset
33
2169
db8baace74d8 Minor Patch for shared libs on Mac OSX by (Bill May <wmay at cisco dot com>)
michael
parents: 2066
diff changeset
34 uint8_t cropTbl[256 + 2 * MAX_NEG_CROP] = {0, };
db8baace74d8 Minor Patch for shared libs on Mac OSX by (Bill May <wmay at cisco dot com>)
michael
parents: 2066
diff changeset
35 uint32_t squareTbl[512] = {0, };
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
36
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
37 const uint8_t ff_zigzag_direct[64] = {
706
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 689
diff changeset
38 0, 1, 8, 16, 9, 2, 3, 10,
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 689
diff changeset
39 17, 24, 32, 25, 18, 11, 4, 5,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
40 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
41 27, 20, 13, 6, 7, 14, 21, 28,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
42 35, 42, 49, 56, 57, 50, 43, 36,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
43 29, 22, 15, 23, 30, 37, 44, 51,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
44 58, 59, 52, 45, 38, 31, 39, 46,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
45 53, 60, 61, 54, 47, 55, 62, 63
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
46 };
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
47
1567
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
48 /* Specific zigzag scan for 248 idct. NOTE that unlike the
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
49 specification, we interleave the fields */
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
50 const uint8_t ff_zigzag248_direct[64] = {
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
51 0, 8, 1, 9, 16, 24, 2, 10,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
52 17, 25, 32, 40, 48, 56, 33, 41,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
53 18, 26, 3, 11, 4, 12, 19, 27,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
54 34, 42, 49, 57, 50, 58, 35, 43,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
55 20, 28, 5, 13, 6, 14, 21, 29,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
56 36, 44, 51, 59, 52, 60, 37, 45,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
57 22, 30, 7, 15, 23, 31, 38, 46,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
58 53, 61, 54, 62, 39, 47, 55, 63,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
59 };
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
60
220
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
61 /* not permutated inverse zigzag_direct + 1 for MMX quantizer */
2169
db8baace74d8 Minor Patch for shared libs on Mac OSX by (Bill May <wmay at cisco dot com>)
michael
parents: 2066
diff changeset
62 uint16_t __align8 inv_zigzag_direct16[64] = {0, };
220
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
63
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
64 const uint8_t ff_alternate_horizontal_scan[64] = {
706
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
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
75 const uint8_t ff_alternate_vertical_scan[64] = {
706
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 */
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
87 const uint32_t inverse[256]={
220
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
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
122 /* Input permutation for the simple_idct_mmx */
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
123 static const uint8_t simple_mmx_permutation[64]={
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
124 0x00, 0x08, 0x04, 0x09, 0x01, 0x0C, 0x05, 0x0D,
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
125 0x10, 0x18, 0x14, 0x19, 0x11, 0x1C, 0x15, 0x1D,
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
126 0x20, 0x28, 0x24, 0x29, 0x21, 0x2C, 0x25, 0x2D,
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
127 0x12, 0x1A, 0x16, 0x1B, 0x13, 0x1E, 0x17, 0x1F,
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
128 0x02, 0x0A, 0x06, 0x0B, 0x03, 0x0E, 0x07, 0x0F,
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
129 0x30, 0x38, 0x34, 0x39, 0x31, 0x3C, 0x35, 0x3D,
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
130 0x22, 0x2A, 0x26, 0x2B, 0x23, 0x2E, 0x27, 0x2F,
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
131 0x32, 0x3A, 0x36, 0x3B, 0x33, 0x3E, 0x37, 0x3F,
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
132 };
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
133
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
134 static int pix_sum_c(uint8_t * pix, int line_size)
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
135 {
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
136 int s, i, j;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
137
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
138 s = 0;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
139 for (i = 0; i < 16; i++) {
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
140 for (j = 0; j < 16; j += 8) {
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
141 s += pix[0];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
142 s += pix[1];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
143 s += pix[2];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
144 s += pix[3];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
145 s += pix[4];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
146 s += pix[5];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
147 s += pix[6];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
148 s += pix[7];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
149 pix += 8;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
150 }
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
151 pix += line_size - 16;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
152 }
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
153 return s;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
154 }
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
155
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
156 static int pix_norm1_c(uint8_t * pix, int line_size)
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
157 {
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
158 int s, i, j;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
159 uint32_t *sq = squareTbl + 256;
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
160
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
161 s = 0;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
162 for (i = 0; i < 16; i++) {
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
163 for (j = 0; j < 16; j += 8) {
997
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
164 #if 0
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
165 s += sq[pix[0]];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
166 s += sq[pix[1]];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
167 s += sq[pix[2]];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
168 s += sq[pix[3]];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
169 s += sq[pix[4]];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
170 s += sq[pix[5]];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
171 s += sq[pix[6]];
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
172 s += sq[pix[7]];
997
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
173 #else
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
174 #if LONG_MAX > 2147483647
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
175 register uint64_t x=*(uint64_t*)pix;
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
176 s += sq[x&0xff];
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
177 s += sq[(x>>8)&0xff];
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
178 s += sq[(x>>16)&0xff];
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
179 s += sq[(x>>24)&0xff];
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
180 s += sq[(x>>32)&0xff];
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
181 s += sq[(x>>40)&0xff];
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
182 s += sq[(x>>48)&0xff];
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
183 s += sq[(x>>56)&0xff];
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
184 #else
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
185 register uint32_t x=*(uint32_t*)pix;
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
186 s += sq[x&0xff];
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
187 s += sq[(x>>8)&0xff];
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
188 s += sq[(x>>16)&0xff];
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
189 s += sq[(x>>24)&0xff];
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
190 x=*(uint32_t*)(pix+4);
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
191 s += sq[x&0xff];
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
192 s += sq[(x>>8)&0xff];
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
193 s += sq[(x>>16)&0xff];
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
194 s += sq[(x>>24)&0xff];
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
195 #endif
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
196 #endif
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
197 pix += 8;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
198 }
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
199 pix += line_size - 16;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
200 }
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
201 return s;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
202 }
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
203
1273
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
204 static void bswap_buf(uint32_t *dst, uint32_t *src, int w){
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
205 int i;
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
206
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
207 for(i=0; i+8<=w; i+=8){
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
208 dst[i+0]= bswap_32(src[i+0]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
209 dst[i+1]= bswap_32(src[i+1]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
210 dst[i+2]= bswap_32(src[i+2]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
211 dst[i+3]= bswap_32(src[i+3]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
212 dst[i+4]= bswap_32(src[i+4]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
213 dst[i+5]= bswap_32(src[i+5]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
214 dst[i+6]= bswap_32(src[i+6]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
215 dst[i+7]= bswap_32(src[i+7]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
216 }
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
217 for(;i<w; i++){
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
218 dst[i+0]= bswap_32(src[i+0]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
219 }
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
220 }
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
221
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
222 static int sse4_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h)
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
223 {
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
224 int s, i;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
225 uint32_t *sq = squareTbl + 256;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
226
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
227 s = 0;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
228 for (i = 0; i < h; i++) {
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
229 s += sq[pix1[0] - pix2[0]];
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
230 s += sq[pix1[1] - pix2[1]];
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
231 s += sq[pix1[2] - pix2[2]];
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
232 s += sq[pix1[3] - pix2[3]];
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
233 pix1 += line_size;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
234 pix2 += line_size;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
235 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
236 return s;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
237 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
238
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
239 static int sse8_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h)
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
240 {
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
241 int s, i;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
242 uint32_t *sq = squareTbl + 256;
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
243
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
244 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
245 for (i = 0; i < h; i++) {
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
246 s += sq[pix1[0] - pix2[0]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
247 s += sq[pix1[1] - pix2[1]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
248 s += sq[pix1[2] - pix2[2]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
249 s += sq[pix1[3] - pix2[3]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
250 s += sq[pix1[4] - pix2[4]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
251 s += sq[pix1[5] - pix2[5]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
252 s += sq[pix1[6] - pix2[6]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
253 s += sq[pix1[7] - pix2[7]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
254 pix1 += line_size;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
255 pix2 += line_size;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
256 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
257 return s;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
258 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
259
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
260 static int sse16_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
884
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 882
diff changeset
261 {
1012
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
262 int s, i;
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
263 uint32_t *sq = squareTbl + 256;
884
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 882
diff changeset
264
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 882
diff changeset
265 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
266 for (i = 0; i < h; i++) {
1012
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
267 s += sq[pix1[ 0] - pix2[ 0]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
268 s += sq[pix1[ 1] - pix2[ 1]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
269 s += sq[pix1[ 2] - pix2[ 2]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
270 s += sq[pix1[ 3] - pix2[ 3]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
271 s += sq[pix1[ 4] - pix2[ 4]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
272 s += sq[pix1[ 5] - pix2[ 5]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
273 s += sq[pix1[ 6] - pix2[ 6]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
274 s += sq[pix1[ 7] - pix2[ 7]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
275 s += sq[pix1[ 8] - pix2[ 8]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
276 s += sq[pix1[ 9] - pix2[ 9]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
277 s += sq[pix1[10] - pix2[10]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
278 s += sq[pix1[11] - pix2[11]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
279 s += sq[pix1[12] - pix2[12]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
280 s += sq[pix1[13] - pix2[13]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
281 s += sq[pix1[14] - pix2[14]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
282 s += sq[pix1[15] - pix2[15]];
997
4dfe15ae0078 sse16 & pix_norm1 optimization patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>) (with some modifications)
michaelni
parents: 996
diff changeset
283
1012
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
284 pix1 += line_size;
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
285 pix2 += line_size;
884
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 882
diff changeset
286 }
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 882
diff changeset
287 return s;
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 882
diff changeset
288 }
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 882
diff changeset
289
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
290
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
291 static inline int w_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int w, int h, int type){
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
292 int s, i, j;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
293 const int dec_count= w==8 ? 3 : 4;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
294 int tmp[16*16];
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
295 #if 0
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
296 int level, ori;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
297 static const int scale[2][2][4][4]={
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
298 {
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
299 {
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
300 //8x8 dec=3
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
301 {268, 239, 239, 213},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
302 { 0, 224, 224, 152},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
303 { 0, 135, 135, 110},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
304 },{
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
305 //16x16 dec=4
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
306 {344, 310, 310, 280},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
307 { 0, 320, 320, 228},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
308 { 0, 175, 175, 136},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
309 { 0, 129, 129, 102},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
310 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
311 },{
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
312 {//FIXME 5/3
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
313 //8x8 dec=3
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
314 {275, 245, 245, 218},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
315 { 0, 230, 230, 156},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
316 { 0, 138, 138, 113},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
317 },{
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
318 //16x16 dec=4
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
319 {352, 317, 317, 286},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
320 { 0, 328, 328, 233},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
321 { 0, 180, 180, 140},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
322 { 0, 132, 132, 105},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
323 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
324 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
325 };
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
326 #endif
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
327
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
328 for (i = 0; i < h; i++) {
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
329 for (j = 0; j < w; j+=4) {
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
330 tmp[16*i+j+0] = (pix1[j+0] - pix2[j+0])<<4;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
331 tmp[16*i+j+1] = (pix1[j+1] - pix2[j+1])<<4;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
332 tmp[16*i+j+2] = (pix1[j+2] - pix2[j+2])<<4;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
333 tmp[16*i+j+3] = (pix1[j+3] - pix2[j+3])<<4;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
334 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
335 pix1 += line_size;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
336 pix2 += line_size;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
337 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
338 ff_spatial_dwt(tmp, w, h, 16, type, dec_count);
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
339
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
340 s=0;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
341 #if 0
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
342 for(level=0; level<dec_count; level++){
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
343 for(ori= level ? 1 : 0; ori<4; ori++){
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
344 int sx= (ori&1) ? 1<<level: 0;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
345 int stride= 16<<(dec_count-level);
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
346 int sy= (ori&2) ? stride>>1 : 0;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
347 int size= 1<<level;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
348
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
349 for(i=0; i<size; i++){
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
350 for(j=0; j<size; j++){
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
351 int v= tmp[sx + sy + i*stride + j] * scale[type][dec_count-3][level][ori];
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
352 s += ABS(v);
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
353 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
354 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
355 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
356 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
357 #endif
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
358 for (i = 0; i < h; i++) {
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
359 for (j = 0; j < w; j+=4) {
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
360 s+= ABS(tmp[16*i+j+0]);
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
361 s+= ABS(tmp[16*i+j+1]);
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
362 s+= ABS(tmp[16*i+j+2]);
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
363 s+= ABS(tmp[16*i+j+3]);
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
364 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
365 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
366 assert(s>=0);
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
367
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
368 return s>>2;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
369 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
370
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
371 static int w53_8_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h){
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
372 return w_c(v, pix1, pix2, line_size, 8, h, 1);
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
373 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
374
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
375 static int w97_8_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h){
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
376 return w_c(v, pix1, pix2, line_size, 8, h, 0);
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
377 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
378
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
379 static int w53_16_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h){
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
380 return w_c(v, pix1, pix2, line_size, 16, h, 1);
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
381 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
382
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
383 static int w97_16_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h){
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
384 return w_c(v, pix1, pix2, line_size, 16, h, 0);
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
385 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
386
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
387 static void get_pixels_c(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
388 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
389 int i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
390
986e461dc072 Initial revision
glantau
parents:
diff changeset
391 /* read the pixels */
986e461dc072 Initial revision
glantau
parents:
diff changeset
392 for(i=0;i<8;i++) {
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
393 block[0] = pixels[0];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
394 block[1] = pixels[1];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
395 block[2] = pixels[2];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
396 block[3] = pixels[3];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
397 block[4] = pixels[4];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
398 block[5] = pixels[5];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
399 block[6] = pixels[6];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
400 block[7] = pixels[7];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
401 pixels += line_size;
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
402 block += 8;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
403 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
404 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
405
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
406 static void diff_pixels_c(DCTELEM *restrict block, const uint8_t *s1,
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
407 const uint8_t *s2, int stride){
324
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
408 int i;
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
409
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
410 /* read the pixels */
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
411 for(i=0;i<8;i++) {
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
412 block[0] = s1[0] - s2[0];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
413 block[1] = s1[1] - s2[1];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
414 block[2] = s1[2] - s2[2];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
415 block[3] = s1[3] - s2[3];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
416 block[4] = s1[4] - s2[4];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
417 block[5] = s1[5] - s2[5];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
418 block[6] = s1[6] - s2[6];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
419 block[7] = s1[7] - s2[7];
324
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
420 s1 += stride;
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
421 s2 += stride;
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
422 block += 8;
324
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
423 }
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
424 }
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
425
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
426
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
427 static void put_pixels_clamped_c(const DCTELEM *block, uint8_t *restrict pixels,
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
428 int line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
429 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
430 int i;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
431 uint8_t *cm = cropTbl + MAX_NEG_CROP;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
432
986e461dc072 Initial revision
glantau
parents:
diff changeset
433 /* read the pixels */
986e461dc072 Initial revision
glantau
parents:
diff changeset
434 for(i=0;i<8;i++) {
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
435 pixels[0] = cm[block[0]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
436 pixels[1] = cm[block[1]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
437 pixels[2] = cm[block[2]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
438 pixels[3] = cm[block[3]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
439 pixels[4] = cm[block[4]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
440 pixels[5] = cm[block[5]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
441 pixels[6] = cm[block[6]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
442 pixels[7] = cm[block[7]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
443
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
444 pixels += line_size;
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
445 block += 8;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
446 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
447 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
448
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
449 static void put_pixels_clamped4_c(const DCTELEM *block, uint8_t *restrict pixels,
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
450 int line_size)
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
451 {
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
452 int i;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
453 uint8_t *cm = cropTbl + MAX_NEG_CROP;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
454
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
455 /* read the pixels */
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
456 for(i=0;i<4;i++) {
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
457 pixels[0] = cm[block[0]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
458 pixels[1] = cm[block[1]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
459 pixels[2] = cm[block[2]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
460 pixels[3] = cm[block[3]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
461
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
462 pixels += line_size;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
463 block += 8;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
464 }
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
465 }
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
466
2257
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
467 static void put_pixels_clamped2_c(const DCTELEM *block, uint8_t *restrict pixels,
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
468 int line_size)
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
469 {
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
470 int i;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
471 uint8_t *cm = cropTbl + MAX_NEG_CROP;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
472
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
473 /* read the pixels */
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
474 for(i=0;i<2;i++) {
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
475 pixels[0] = cm[block[0]];
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
476 pixels[1] = cm[block[1]];
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
477
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
478 pixels += line_size;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
479 block += 8;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
480 }
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
481 }
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
482
1984
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
483 static void put_signed_pixels_clamped_c(const DCTELEM *block,
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
484 uint8_t *restrict pixels,
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
485 int line_size)
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
486 {
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
487 int i, j;
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
488
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
489 for (i = 0; i < 8; i++) {
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
490 for (j = 0; j < 8; j++) {
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
491 if (*block < -128)
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
492 *pixels = 0;
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
493 else if (*block > 127)
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
494 *pixels = 255;
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
495 else
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
496 *pixels = (uint8_t)(*block + 128);
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
497 block++;
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
498 pixels++;
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
499 }
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
500 pixels += (line_size - 8);
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
501 }
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
502 }
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
503
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
504 static void add_pixels_clamped_c(const DCTELEM *block, uint8_t *restrict pixels,
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
505 int line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
506 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
507 int i;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
508 uint8_t *cm = cropTbl + MAX_NEG_CROP;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
509
986e461dc072 Initial revision
glantau
parents:
diff changeset
510 /* read the pixels */
986e461dc072 Initial revision
glantau
parents:
diff changeset
511 for(i=0;i<8;i++) {
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
512 pixels[0] = cm[pixels[0] + block[0]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
513 pixels[1] = cm[pixels[1] + block[1]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
514 pixels[2] = cm[pixels[2] + block[2]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
515 pixels[3] = cm[pixels[3] + block[3]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
516 pixels[4] = cm[pixels[4] + block[4]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
517 pixels[5] = cm[pixels[5] + block[5]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
518 pixels[6] = cm[pixels[6] + block[6]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
519 pixels[7] = cm[pixels[7] + block[7]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
520 pixels += line_size;
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
521 block += 8;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
522 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
523 }
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
524
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
525 static void add_pixels_clamped4_c(const DCTELEM *block, uint8_t *restrict pixels,
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
526 int line_size)
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
527 {
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
528 int i;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
529 uint8_t *cm = cropTbl + MAX_NEG_CROP;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
530
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
531 /* read the pixels */
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
532 for(i=0;i<4;i++) {
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
533 pixels[0] = cm[pixels[0] + block[0]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
534 pixels[1] = cm[pixels[1] + block[1]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
535 pixels[2] = cm[pixels[2] + block[2]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
536 pixels[3] = cm[pixels[3] + block[3]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
537 pixels += line_size;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
538 block += 8;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
539 }
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
540 }
2257
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
541
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
542 static void add_pixels_clamped2_c(const DCTELEM *block, uint8_t *restrict pixels,
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
543 int line_size)
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
544 {
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
545 int i;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
546 uint8_t *cm = cropTbl + MAX_NEG_CROP;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
547
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
548 /* read the pixels */
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
549 for(i=0;i<2;i++) {
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
550 pixels[0] = cm[pixels[0] + block[0]];
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
551 pixels[1] = cm[pixels[1] + block[1]];
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
552 pixels += line_size;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
553 block += 8;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
554 }
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
555 }
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
556 #if 0
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
557
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
558 #define PIXOP2(OPNAME, OP) \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
559 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
560 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
561 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
562 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
563 OP(*((uint64_t*)block), LD64(pixels));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
564 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
565 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
566 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
567 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
568 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
569 static void OPNAME ## _no_rnd_pixels_x2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
570 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
571 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
572 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
573 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
574 const uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
575 OP(*((uint64_t*)block), (a&b) + (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
576 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
577 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
578 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
579 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
580 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
581 static void OPNAME ## _pixels_x2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
582 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
583 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
584 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
585 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
586 const uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
587 OP(*((uint64_t*)block), (a|b) - (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
588 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
589 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
590 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
591 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
592 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
593 static void OPNAME ## _no_rnd_pixels_y2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
594 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
595 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
596 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
597 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
598 const uint64_t b= LD64(pixels+line_size);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
599 OP(*((uint64_t*)block), (a&b) + (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
600 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
601 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
602 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
603 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
604 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
605 static void OPNAME ## _pixels_y2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
606 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
607 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
608 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
609 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
610 const uint64_t b= LD64(pixels+line_size);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
611 OP(*((uint64_t*)block), (a|b) - (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
612 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
613 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
614 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
615 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
616 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
617 static void OPNAME ## _pixels_xy2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
618 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
619 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
620 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
621 const uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
622 uint64_t l0= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
623 + (b&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
624 + 0x0202020202020202ULL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
625 uint64_t h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
626 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
627 uint64_t l1,h1;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
628 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
629 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
630 for(i=0; i<h; i+=2){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
631 uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
632 uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
633 l1= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
634 + (b&0x0303030303030303ULL);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
635 h1= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
636 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
637 OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
638 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
639 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
640 a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
641 b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
642 l0= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
643 + (b&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
644 + 0x0202020202020202ULL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
645 h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
646 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
647 OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
648 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
649 block +=line_size;\
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 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
653 static void OPNAME ## _no_rnd_pixels_xy2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
654 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
655 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
656 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
657 const uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
658 uint64_t l0= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
659 + (b&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
660 + 0x0101010101010101ULL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
661 uint64_t h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
662 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
663 uint64_t l1,h1;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
664 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
665 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
666 for(i=0; i<h; i+=2){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
667 uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
668 uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
669 l1= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
670 + (b&0x0303030303030303ULL);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
671 h1= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
672 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
673 OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
674 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
675 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
676 a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
677 b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
678 l0= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
679 + (b&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
680 + 0x0101010101010101ULL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
681 h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
682 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
683 OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
684 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
685 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
686 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
687 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
688 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
689 CALL_2X_PIXELS(OPNAME ## _pixels16_c , OPNAME ## _pixels_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
690 CALL_2X_PIXELS(OPNAME ## _pixels16_x2_c , OPNAME ## _pixels_x2_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
691 CALL_2X_PIXELS(OPNAME ## _pixels16_y2_c , OPNAME ## _pixels_y2_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
692 CALL_2X_PIXELS(OPNAME ## _pixels16_xy2_c, OPNAME ## _pixels_xy2_c, 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
693 CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_x2_c , OPNAME ## _no_rnd_pixels_x2_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
694 CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_y2_c , OPNAME ## _no_rnd_pixels_y2_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
695 CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_xy2_c, OPNAME ## _no_rnd_pixels_xy2_c, 8)
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
696
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
697 #define op_avg(a, b) a = ( ((a)|(b)) - ((((a)^(b))&0xFEFEFEFEFEFEFEFEULL)>>1) )
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
698 #else // 64 bit variant
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
699
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
700 #define PIXOP2(OPNAME, OP) \
1267
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
701 static void OPNAME ## _pixels2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
702 int i;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
703 for(i=0; i<h; i++){\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
704 OP(*((uint16_t*)(block )), LD16(pixels ));\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
705 pixels+=line_size;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
706 block +=line_size;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
707 }\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
708 }\
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
709 static void OPNAME ## _pixels4_c(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
710 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
711 for(i=0; i<h; i++){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
712 OP(*((uint32_t*)(block )), LD32(pixels ));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
713 pixels+=line_size;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
714 block +=line_size;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
715 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
716 }\
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
717 static void OPNAME ## _pixels8_c(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
718 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
719 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
720 OP(*((uint32_t*)(block )), LD32(pixels ));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
721 OP(*((uint32_t*)(block+4)), LD32(pixels+4));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
722 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
723 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
724 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
725 }\
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
726 static inline void OPNAME ## _no_rnd_pixels8_c(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
727 OPNAME ## _pixels8_c(block, pixels, line_size, h);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
728 }\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
729 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
730 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
731 int src_stride1, int src_stride2, int h){\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
732 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
733 for(i=0; i<h; i++){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
734 uint32_t a,b;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
735 a= LD32(&src1[i*src_stride1 ]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
736 b= LD32(&src2[i*src_stride2 ]);\
1264
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
737 OP(*((uint32_t*)&dst[i*dst_stride ]), no_rnd_avg32(a, b));\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
738 a= LD32(&src1[i*src_stride1+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
739 b= LD32(&src2[i*src_stride2+4]);\
1264
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
740 OP(*((uint32_t*)&dst[i*dst_stride+4]), no_rnd_avg32(a, b));\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
741 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
742 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
743 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
744 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
745 int src_stride1, int src_stride2, int h){\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
746 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
747 for(i=0; i<h; i++){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
748 uint32_t a,b;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
749 a= LD32(&src1[i*src_stride1 ]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
750 b= LD32(&src2[i*src_stride2 ]);\
1264
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
751 OP(*((uint32_t*)&dst[i*dst_stride ]), rnd_avg32(a, b));\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
752 a= LD32(&src1[i*src_stride1+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
753 b= LD32(&src2[i*src_stride2+4]);\
1264
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
754 OP(*((uint32_t*)&dst[i*dst_stride+4]), rnd_avg32(a, b));\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
755 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
756 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
757 \
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
758 static inline void OPNAME ## _pixels4_l2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
759 int src_stride1, int src_stride2, int h){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
760 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
761 for(i=0; i<h; i++){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
762 uint32_t a,b;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
763 a= LD32(&src1[i*src_stride1 ]);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
764 b= LD32(&src2[i*src_stride2 ]);\
1264
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
765 OP(*((uint32_t*)&dst[i*dst_stride ]), rnd_avg32(a, b));\
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
766 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
767 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
768 \
1267
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
769 static inline void OPNAME ## _pixels2_l2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
770 int src_stride1, int src_stride2, int h){\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
771 int i;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
772 for(i=0; i<h; i++){\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
773 uint32_t a,b;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
774 a= LD16(&src1[i*src_stride1 ]);\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
775 b= LD16(&src2[i*src_stride2 ]);\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
776 OP(*((uint16_t*)&dst[i*dst_stride ]), rnd_avg32(a, b));\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
777 }\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
778 }\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
779 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
780 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
781 int src_stride1, int src_stride2, int h){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
782 OPNAME ## _pixels8_l2(dst , src1 , src2 , dst_stride, src_stride1, src_stride2, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
783 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
784 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
785 \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
786 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
787 int src_stride1, int src_stride2, int h){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
788 OPNAME ## _no_rnd_pixels8_l2(dst , src1 , src2 , dst_stride, src_stride1, src_stride2, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
789 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
790 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
791 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
792 static inline void OPNAME ## _no_rnd_pixels8_x2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
793 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
794 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
795 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
796 static inline void OPNAME ## _pixels8_x2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
797 OPNAME ## _pixels8_l2(block, pixels, pixels+1, line_size, line_size, line_size, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
798 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
799 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
800 static inline void OPNAME ## _no_rnd_pixels8_y2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
801 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
802 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
803 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
804 static inline void OPNAME ## _pixels8_y2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
805 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
806 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
807 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
808 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
809 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
810 int i;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
811 for(i=0; i<h; i++){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
812 uint32_t a, b, c, d, l0, l1, h0, h1;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
813 a= LD32(&src1[i*src_stride1]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
814 b= LD32(&src2[i*src_stride2]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
815 c= LD32(&src3[i*src_stride3]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
816 d= LD32(&src4[i*src_stride4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
817 l0= (a&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
818 + (b&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
819 + 0x02020202UL;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
820 h0= ((a&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
821 + ((b&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
822 l1= (c&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
823 + (d&0x03030303UL);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
824 h1= ((c&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
825 + ((d&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
826 OP(*((uint32_t*)&dst[i*dst_stride]), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
827 a= LD32(&src1[i*src_stride1+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
828 b= LD32(&src2[i*src_stride2+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
829 c= LD32(&src3[i*src_stride3+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
830 d= LD32(&src4[i*src_stride4+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
831 l0= (a&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
832 + (b&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
833 + 0x02020202UL;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
834 h0= ((a&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
835 + ((b&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
836 l1= (c&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
837 + (d&0x03030303UL);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
838 h1= ((c&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
839 + ((d&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
840 OP(*((uint32_t*)&dst[i*dst_stride+4]), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
841 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
842 }\
1267
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
843 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
844 static inline void OPNAME ## _pixels4_x2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
845 OPNAME ## _pixels4_l2(block, pixels, pixels+1, line_size, line_size, line_size, h);\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
846 }\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
847 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
848 static inline void OPNAME ## _pixels4_y2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
849 OPNAME ## _pixels4_l2(block, pixels, pixels+line_size, line_size, line_size, line_size, h);\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
850 }\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
851 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
852 static inline void OPNAME ## _pixels2_x2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
853 OPNAME ## _pixels2_l2(block, pixels, pixels+1, line_size, line_size, line_size, h);\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
854 }\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
855 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
856 static inline void OPNAME ## _pixels2_y2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
857 OPNAME ## _pixels2_l2(block, pixels, pixels+line_size, line_size, line_size, line_size, h);\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
858 }\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
859 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
860 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
861 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
862 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
863 for(i=0; i<h; i++){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
864 uint32_t a, b, c, d, l0, l1, h0, h1;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
865 a= LD32(&src1[i*src_stride1]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
866 b= LD32(&src2[i*src_stride2]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
867 c= LD32(&src3[i*src_stride3]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
868 d= LD32(&src4[i*src_stride4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
869 l0= (a&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
870 + (b&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
871 + 0x01010101UL;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
872 h0= ((a&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
873 + ((b&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
874 l1= (c&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
875 + (d&0x03030303UL);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
876 h1= ((c&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
877 + ((d&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
878 OP(*((uint32_t*)&dst[i*dst_stride]), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
879 a= LD32(&src1[i*src_stride1+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
880 b= LD32(&src2[i*src_stride2+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
881 c= LD32(&src3[i*src_stride3+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
882 d= LD32(&src4[i*src_stride4+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
883 l0= (a&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
884 + (b&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
885 + 0x01010101UL;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
886 h0= ((a&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
887 + ((b&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
888 l1= (c&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
889 + (d&0x03030303UL);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
890 h1= ((c&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
891 + ((d&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
892 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
893 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
894 }\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
895 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
896 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
897 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
898 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
899 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
900 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
901 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
902 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
903 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
904 }\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
905 \
1267
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
906 static inline void OPNAME ## _pixels2_xy2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
907 {\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
908 int i, a0, b0, a1, b1;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
909 a0= pixels[0];\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
910 b0= pixels[1] + 2;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
911 a0 += b0;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
912 b0 += pixels[2];\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
913 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
914 pixels+=line_size;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
915 for(i=0; i<h; i+=2){\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
916 a1= pixels[0];\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
917 b1= pixels[1];\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
918 a1 += b1;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
919 b1 += pixels[2];\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
920 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
921 block[0]= (a1+a0)>>2; /* FIXME non put */\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
922 block[1]= (b1+b0)>>2;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
923 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
924 pixels+=line_size;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
925 block +=line_size;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
926 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
927 a0= pixels[0];\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
928 b0= pixels[1] + 2;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
929 a0 += b0;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
930 b0 += pixels[2];\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
931 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
932 block[0]= (a1+a0)>>2;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
933 block[1]= (b1+b0)>>2;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
934 pixels+=line_size;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
935 block +=line_size;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
936 }\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
937 }\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
938 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
939 static inline void OPNAME ## _pixels4_xy2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
940 {\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
941 int i;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
942 const uint32_t a= LD32(pixels );\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
943 const uint32_t b= LD32(pixels+1);\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
944 uint32_t l0= (a&0x03030303UL)\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
945 + (b&0x03030303UL)\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
946 + 0x02020202UL;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
947 uint32_t h0= ((a&0xFCFCFCFCUL)>>2)\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
948 + ((b&0xFCFCFCFCUL)>>2);\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
949 uint32_t l1,h1;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
950 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
951 pixels+=line_size;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
952 for(i=0; i<h; i+=2){\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
953 uint32_t a= LD32(pixels );\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
954 uint32_t b= LD32(pixels+1);\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
955 l1= (a&0x03030303UL)\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
956 + (b&0x03030303UL);\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
957 h1= ((a&0xFCFCFCFCUL)>>2)\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
958 + ((b&0xFCFCFCFCUL)>>2);\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
959 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
960 pixels+=line_size;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
961 block +=line_size;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
962 a= LD32(pixels );\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
963 b= LD32(pixels+1);\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
964 l0= (a&0x03030303UL)\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
965 + (b&0x03030303UL)\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
966 + 0x02020202UL;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
967 h0= ((a&0xFCFCFCFCUL)>>2)\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
968 + ((b&0xFCFCFCFCUL)>>2);\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
969 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
970 pixels+=line_size;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
971 block +=line_size;\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
972 }\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
973 }\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
974 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
975 static inline void OPNAME ## _pixels8_xy2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
976 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
977 int j;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
978 for(j=0; j<2; j++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
979 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
980 const uint32_t a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
981 const uint32_t b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
982 uint32_t l0= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
983 + (b&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
984 + 0x02020202UL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
985 uint32_t h0= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
986 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
987 uint32_t l1,h1;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
988 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
989 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
990 for(i=0; i<h; i+=2){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
991 uint32_t a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
992 uint32_t b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
993 l1= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
994 + (b&0x03030303UL);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
995 h1= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
996 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
997 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
998 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
999 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1000 a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1001 b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1002 l0= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1003 + (b&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1004 + 0x02020202UL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1005 h0= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1006 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1007 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1008 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1009 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1010 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1011 pixels+=4-line_size*(h+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1012 block +=4-line_size*h;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1013 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1014 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1015 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1016 static inline void OPNAME ## _no_rnd_pixels8_xy2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1017 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1018 int j;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1019 for(j=0; j<2; j++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1020 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1021 const uint32_t a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1022 const uint32_t b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1023 uint32_t l0= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1024 + (b&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1025 + 0x01010101UL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1026 uint32_t h0= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1027 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1028 uint32_t l1,h1;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1029 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1030 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1031 for(i=0; i<h; i+=2){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1032 uint32_t a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1033 uint32_t b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1034 l1= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1035 + (b&0x03030303UL);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1036 h1= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1037 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1038 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1039 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1040 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1041 a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1042 b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1043 l0= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1044 + (b&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1045 + 0x01010101UL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1046 h0= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1047 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1048 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1049 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1050 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1051 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1052 pixels+=4-line_size*(h+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1053 block +=4-line_size*h;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1054 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1055 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1056 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1057 CALL_2X_PIXELS(OPNAME ## _pixels16_c , OPNAME ## _pixels8_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1058 CALL_2X_PIXELS(OPNAME ## _pixels16_x2_c , OPNAME ## _pixels8_x2_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1059 CALL_2X_PIXELS(OPNAME ## _pixels16_y2_c , OPNAME ## _pixels8_y2_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1060 CALL_2X_PIXELS(OPNAME ## _pixels16_xy2_c, OPNAME ## _pixels8_xy2_c, 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1061 CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_c , OPNAME ## _pixels8_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1062 CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_x2_c , OPNAME ## _no_rnd_pixels8_x2_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1063 CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_y2_c , OPNAME ## _no_rnd_pixels8_y2_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1064 CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_xy2_c, OPNAME ## _no_rnd_pixels8_xy2_c, 8)\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1065
1264
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
1066 #define op_avg(a, b) a = rnd_avg32(a, b)
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1067 #endif
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1068 #define op_put(a, b) a = b
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1069
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1070 PIXOP2(avg, op_avg)
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1071 PIXOP2(put, op_put)
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1072 #undef op_avg
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1073 #undef op_put
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1074
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1075 #define avg2(a,b) ((a+b+1)>>1)
986e461dc072 Initial revision
glantau
parents:
diff changeset
1076 #define avg4(a,b,c,d) ((a+b+c+d+2)>>2)
986e461dc072 Initial revision
glantau
parents:
diff changeset
1077
1864
9a73e6dc5d68 fix motion compensation with (x+1/2,y+1/2) MVs
michael
parents: 1784
diff changeset
1078 static void put_no_rnd_pixels16_l2_c(uint8_t *dst, const uint8_t *a, const uint8_t *b, int stride, int h){
9a73e6dc5d68 fix motion compensation with (x+1/2,y+1/2) MVs
michael
parents: 1784
diff changeset
1079 put_no_rnd_pixels16_l2(dst, a, b, stride, stride, stride, h);
9a73e6dc5d68 fix motion compensation with (x+1/2,y+1/2) MVs
michael
parents: 1784
diff changeset
1080 }
9a73e6dc5d68 fix motion compensation with (x+1/2,y+1/2) MVs
michael
parents: 1784
diff changeset
1081
9a73e6dc5d68 fix motion compensation with (x+1/2,y+1/2) MVs
michael
parents: 1784
diff changeset
1082 static void put_no_rnd_pixels8_l2_c(uint8_t *dst, const uint8_t *a, const uint8_t *b, int stride, int h){
9a73e6dc5d68 fix motion compensation with (x+1/2,y+1/2) MVs
michael
parents: 1784
diff changeset
1083 put_no_rnd_pixels8_l2(dst, a, b, stride, stride, stride, h);
9a73e6dc5d68 fix motion compensation with (x+1/2,y+1/2) MVs
michael
parents: 1784
diff changeset
1084 }
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1085
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1086 static void gmc1_c(uint8_t *dst, uint8_t *src, int stride, int h, int x16, int y16, int rounder)
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1087 {
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1088 const int A=(16-x16)*(16-y16);
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1089 const int B=( x16)*(16-y16);
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1090 const int C=(16-x16)*( y16);
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1091 const int D=( x16)*( y16);
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1092 int i;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1093
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1094 for(i=0; i<h; i++)
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1095 {
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1096 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
1097 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
1098 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
1099 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
1100 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
1101 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
1102 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
1103 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
1104 dst+= stride;
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1105 src+= stride;
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1106 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1107 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1108
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1109 static void gmc_c(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1110 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
1111 {
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1112 int y, vx, vy;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1113 const int s= 1<<shift;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1114
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1115 width--;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1116 height--;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1117
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1118 for(y=0; y<h; y++){
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1119 int x;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1120
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1121 vx= ox;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1122 vy= oy;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1123 for(x=0; x<8; x++){ //XXX FIXME optimize
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1124 int src_x, src_y, frac_x, frac_y, index;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1125
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1126 src_x= vx>>16;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1127 src_y= vy>>16;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1128 frac_x= src_x&(s-1);
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1129 frac_y= src_y&(s-1);
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1130 src_x>>=shift;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1131 src_y>>=shift;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1132
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1133 if((unsigned)src_x < width){
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1134 if((unsigned)src_y < height){
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1135 index= src_x + src_y*stride;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1136 dst[y*stride + x]= ( ( src[index ]*(s-frac_x)
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1137 + src[index +1]* frac_x )*(s-frac_y)
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1138 + ( src[index+stride ]*(s-frac_x)
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1139 + src[index+stride+1]* frac_x )* frac_y
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1140 + r)>>(shift*2);
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1141 }else{
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1142 index= src_x + clip(src_y, 0, height)*stride;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1143 dst[y*stride + x]= ( ( src[index ]*(s-frac_x)
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1144 + src[index +1]* frac_x )*s
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1145 + r)>>(shift*2);
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1146 }
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1147 }else{
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1148 if((unsigned)src_y < height){
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1149 index= clip(src_x, 0, width) + src_y*stride;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1150 dst[y*stride + x]= ( ( src[index ]*(s-frac_y)
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1151 + src[index+stride ]* frac_y )*s
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1152 + r)>>(shift*2);
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1153 }else{
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1154 index= clip(src_x, 0, width) + clip(src_y, 0, height)*stride;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1155 dst[y*stride + x]= src[index ];
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1156 }
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1157 }
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1158
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1159 vx+= dxx;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1160 vy+= dyx;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1161 }
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1162 ox += dxy;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1163 oy += dyy;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1164 }
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1165 }
1267
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1166
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1167 static inline void put_tpel_pixels_mc00_c(uint8_t *dst, const uint8_t *src, int stride, int width, int height){
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1168 switch(width){
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1169 case 2: put_pixels2_c (dst, src, stride, height); break;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1170 case 4: put_pixels4_c (dst, src, stride, height); break;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1171 case 8: put_pixels8_c (dst, src, stride, height); break;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1172 case 16:put_pixels16_c(dst, src, stride, height); break;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1173 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1174 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1175
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1176 static inline void put_tpel_pixels_mc10_c(uint8_t *dst, const uint8_t *src, int stride, int width, int height){
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1177 int i,j;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1178 for (i=0; i < height; i++) {
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1179 for (j=0; j < width; j++) {
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1180 dst[j] = (683*(2*src[j] + src[j+1] + 1)) >> 11;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1181 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1182 src += stride;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1183 dst += stride;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1184 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1185 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1186
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1187 static inline void put_tpel_pixels_mc20_c(uint8_t *dst, const uint8_t *src, int stride, int width, int height){
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1188 int i,j;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1189 for (i=0; i < height; i++) {
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1190 for (j=0; j < width; j++) {
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1191 dst[j] = (683*(src[j] + 2*src[j+1] + 1)) >> 11;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1192 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1193 src += stride;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1194 dst += stride;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1195 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1196 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1197
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1198 static inline void put_tpel_pixels_mc01_c(uint8_t *dst, const uint8_t *src, int stride, int width, int height){
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1199 int i,j;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1200 for (i=0; i < height; i++) {
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1201 for (j=0; j < width; j++) {
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1202 dst[j] = (683*(2*src[j] + src[j+stride] + 1)) >> 11;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1203 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1204 src += stride;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1205 dst += stride;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1206 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1207 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1208
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1209 static inline void put_tpel_pixels_mc11_c(uint8_t *dst, const uint8_t *src, int stride, int width, int height){
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1210 int i,j;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1211 for (i=0; i < height; i++) {
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1212 for (j=0; j < width; j++) {
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1213 dst[j] = (2731*(4*src[j] + 3*src[j+1] + 3*src[j+stride] + 2*src[j+stride+1] + 6)) >> 15;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1214 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1215 src += stride;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1216 dst += stride;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1217 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1218 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1219
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1220 static inline void put_tpel_pixels_mc12_c(uint8_t *dst, const uint8_t *src, int stride, int width, int height){
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1221 int i,j;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1222 for (i=0; i < height; i++) {
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1223 for (j=0; j < width; j++) {
1329
2114c1e206e8 Fix third pixel motion compensation
tmmm
parents: 1324
diff changeset
1224 dst[j] = (2731*(3*src[j] + 2*src[j+1] + 4*src[j+stride] + 3*src[j+stride+1] + 6)) >> 15;
1267
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1225 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1226 src += stride;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1227 dst += stride;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1228 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1229 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1230
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1231 static inline void put_tpel_pixels_mc02_c(uint8_t *dst, const uint8_t *src, int stride, int width, int height){
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1232 int i,j;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1233 for (i=0; i < height; i++) {
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1234 for (j=0; j < width; j++) {
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1235 dst[j] = (683*(src[j] + 2*src[j+stride] + 1)) >> 11;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1236 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1237 src += stride;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1238 dst += stride;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1239 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1240 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1241
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1242 static inline void put_tpel_pixels_mc21_c(uint8_t *dst, const uint8_t *src, int stride, int width, int height){
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1243 int i,j;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1244 for (i=0; i < height; i++) {
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1245 for (j=0; j < width; j++) {
1329
2114c1e206e8 Fix third pixel motion compensation
tmmm
parents: 1324
diff changeset
1246 dst[j] = (2731*(3*src[j] + 4*src[j+1] + 2*src[j+stride] + 3*src[j+stride+1] + 6)) >> 15;
1267
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1247 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1248 src += stride;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1249 dst += stride;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1250 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1251 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1252
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1253 static inline void put_tpel_pixels_mc22_c(uint8_t *dst, const uint8_t *src, int stride, int width, int height){
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1254 int i,j;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1255 for (i=0; i < height; i++) {
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1256 for (j=0; j < width; j++) {
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1257 dst[j] = (2731*(2*src[j] + 3*src[j+1] + 3*src[j+stride] + 4*src[j+stride+1] + 6)) >> 15;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1258 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1259 src += stride;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1260 dst += stride;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1261 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1262 }
1319
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1263
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1264 static inline void avg_tpel_pixels_mc00_c(uint8_t *dst, const uint8_t *src, int stride, int width, int height){
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1265 switch(width){
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1266 case 2: avg_pixels2_c (dst, src, stride, height); break;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1267 case 4: avg_pixels4_c (dst, src, stride, height); break;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1268 case 8: avg_pixels8_c (dst, src, stride, height); break;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1269 case 16:avg_pixels16_c(dst, src, stride, height); break;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1270 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1271 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1272
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1273 static inline void avg_tpel_pixels_mc10_c(uint8_t *dst, const uint8_t *src, int stride, int width, int height){
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1274 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1275 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1276 for (j=0; j < width; j++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1277 dst[j] = (dst[j] + ((683*(2*src[j] + src[j+1] + 1)) >> 11) + 1) >> 1;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1278 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1279 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1280 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1281 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1282 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1283
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1284 static inline void avg_tpel_pixels_mc20_c(uint8_t *dst, const uint8_t *src, int stride, int width, int height){
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1285 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1286 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1287 for (j=0; j < width; j++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1288 dst[j] = (dst[j] + ((683*(src[j] + 2*src[j+1] + 1)) >> 11) + 1) >> 1;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1289 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1290 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1291 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1292 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1293 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1294
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1295 static inline void avg_tpel_pixels_mc01_c(uint8_t *dst, const uint8_t *src, int stride, int width, int height){
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1296 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1297 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1298 for (j=0; j < width; j++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1299 dst[j] = (dst[j] + ((683*(2*src[j] + src[j+stride] + 1)) >> 11) + 1) >> 1;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1300 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1301 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1302 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1303 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1304 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1305
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1306 static inline void avg_tpel_pixels_mc11_c(uint8_t *dst, const uint8_t *src, int stride, int width, int height){
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1307 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1308 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1309 for (j=0; j < width; j++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1310 dst[j] = (dst[j] + ((2731*(4*src[j] + 3*src[j+1] + 3*src[j+stride] + 2*src[j+stride+1] + 6)) >> 15) + 1) >> 1;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1311 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1312 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1313 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1314 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1315 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1316
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1317 static inline void avg_tpel_pixels_mc12_c(uint8_t *dst, const uint8_t *src, int stride, int width, int height){
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1318 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1319 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1320 for (j=0; j < width; j++) {
1329
2114c1e206e8 Fix third pixel motion compensation
tmmm
parents: 1324
diff changeset
1321 dst[j] = (dst[j] + ((2731*(3*src[j] + 2*src[j+1] + 4*src[j+stride] + 3*src[j+stride+1] + 6)) >> 15) + 1) >> 1;
1319
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1322 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1323 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1324 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1325 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1326 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1327
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1328 static inline void avg_tpel_pixels_mc02_c(uint8_t *dst, const uint8_t *src, int stride, int width, int height){
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1329 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1330 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1331 for (j=0; j < width; j++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1332 dst[j] = (dst[j] + ((683*(src[j] + 2*src[j+stride] + 1)) >> 11) + 1) >> 1;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1333 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1334 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1335 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1336 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1337 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1338
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1339 static inline void avg_tpel_pixels_mc21_c(uint8_t *dst, const uint8_t *src, int stride, int width, int height){
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1340 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1341 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1342 for (j=0; j < width; j++) {
1329
2114c1e206e8 Fix third pixel motion compensation
tmmm
parents: 1324
diff changeset
1343 dst[j] = (dst[j] + ((2731*(3*src[j] + 4*src[j+1] + 2*src[j+stride] + 3*src[j+stride+1] + 6)) >> 15) + 1) >> 1;
1319
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1344 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1345 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1346 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1347 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1348 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1349
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1350 static inline void avg_tpel_pixels_mc22_c(uint8_t *dst, const uint8_t *src, int stride, int width, int height){
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1351 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1352 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1353 for (j=0; j < width; j++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1354 dst[j] = (dst[j] + ((2731*(2*src[j] + 3*src[j+1] + 3*src[j+stride] + 4*src[j+stride+1] + 6)) >> 15) + 1) >> 1;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1355 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1356 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1357 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1358 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1359 }
1267
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1360 #if 0
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1361 #define TPEL_WIDTH(width)\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1362 static void put_tpel_pixels ## width ## _mc00_c(uint8_t *dst, const uint8_t *src, int stride, int height){\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1363 void put_tpel_pixels_mc00_c(dst, src, stride, width, height);}\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1364 static void put_tpel_pixels ## width ## _mc10_c(uint8_t *dst, const uint8_t *src, int stride, int height){\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1365 void put_tpel_pixels_mc10_c(dst, src, stride, width, height);}\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1366 static void put_tpel_pixels ## width ## _mc20_c(uint8_t *dst, const uint8_t *src, int stride, int height){\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1367 void put_tpel_pixels_mc20_c(dst, src, stride, width, height);}\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1368 static void put_tpel_pixels ## width ## _mc01_c(uint8_t *dst, const uint8_t *src, int stride, int height){\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1369 void put_tpel_pixels_mc01_c(dst, src, stride, width, height);}\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1370 static void put_tpel_pixels ## width ## _mc11_c(uint8_t *dst, const uint8_t *src, int stride, int height){\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1371 void put_tpel_pixels_mc11_c(dst, src, stride, width, height);}\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1372 static void put_tpel_pixels ## width ## _mc21_c(uint8_t *dst, const uint8_t *src, int stride, int height){\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1373 void put_tpel_pixels_mc21_c(dst, src, stride, width, height);}\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1374 static void put_tpel_pixels ## width ## _mc02_c(uint8_t *dst, const uint8_t *src, int stride, int height){\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1375 void put_tpel_pixels_mc02_c(dst, src, stride, width, height);}\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1376 static void put_tpel_pixels ## width ## _mc12_c(uint8_t *dst, const uint8_t *src, int stride, int height){\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1377 void put_tpel_pixels_mc12_c(dst, src, stride, width, height);}\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1378 static void put_tpel_pixels ## width ## _mc22_c(uint8_t *dst, const uint8_t *src, int stride, int height){\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1379 void put_tpel_pixels_mc22_c(dst, src, stride, width, height);}
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1380 #endif
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1381
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1382 #define H264_CHROMA_MC(OPNAME, OP)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1383 static void OPNAME ## h264_chroma_mc2_c(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1384 const int A=(8-x)*(8-y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1385 const int B=( x)*(8-y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1386 const int C=(8-x)*( y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1387 const int D=( x)*( y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1388 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1389 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1390 assert(x<8 && y<8 && x>=0 && y>=0);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1391 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1392 for(i=0; i<h; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1393 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1394 OP(dst[0], (A*src[0] + B*src[1] + C*src[stride+0] + D*src[stride+1]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1395 OP(dst[1], (A*src[1] + B*src[2] + C*src[stride+1] + D*src[stride+2]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1396 dst+= stride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1397 src+= stride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1398 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1399 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1400 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1401 static void OPNAME ## h264_chroma_mc4_c(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1402 const int A=(8-x)*(8-y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1403 const int B=( x)*(8-y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1404 const int C=(8-x)*( y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1405 const int D=( x)*( y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1406 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1407 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1408 assert(x<8 && y<8 && x>=0 && y>=0);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1409 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1410 for(i=0; i<h; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1411 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1412 OP(dst[0], (A*src[0] + B*src[1] + C*src[stride+0] + D*src[stride+1]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1413 OP(dst[1], (A*src[1] + B*src[2] + C*src[stride+1] + D*src[stride+2]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1414 OP(dst[2], (A*src[2] + B*src[3] + C*src[stride+2] + D*src[stride+3]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1415 OP(dst[3], (A*src[3] + B*src[4] + C*src[stride+3] + D*src[stride+4]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1416 dst+= stride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1417 src+= stride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1418 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1419 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1420 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1421 static void OPNAME ## h264_chroma_mc8_c(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1422 const int A=(8-x)*(8-y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1423 const int B=( x)*(8-y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1424 const int C=(8-x)*( y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1425 const int D=( x)*( y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1426 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1427 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1428 assert(x<8 && y<8 && x>=0 && y>=0);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1429 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1430 for(i=0; i<h; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1431 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1432 OP(dst[0], (A*src[0] + B*src[1] + C*src[stride+0] + D*src[stride+1]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1433 OP(dst[1], (A*src[1] + B*src[2] + C*src[stride+1] + D*src[stride+2]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1434 OP(dst[2], (A*src[2] + B*src[3] + C*src[stride+2] + D*src[stride+3]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1435 OP(dst[3], (A*src[3] + B*src[4] + C*src[stride+3] + D*src[stride+4]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1436 OP(dst[4], (A*src[4] + B*src[5] + C*src[stride+4] + D*src[stride+5]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1437 OP(dst[5], (A*src[5] + B*src[6] + C*src[stride+5] + D*src[stride+6]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1438 OP(dst[6], (A*src[6] + B*src[7] + C*src[stride+6] + D*src[stride+7]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1439 OP(dst[7], (A*src[7] + B*src[8] + C*src[stride+7] + D*src[stride+8]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1440 dst+= stride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1441 src+= stride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1442 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1443 }
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1444
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1445 #define op_avg(a, b) a = (((a)+(((b) + 32)>>6)+1)>>1)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1446 #define op_put(a, b) a = (((b) + 32)>>6)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1447
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1448 H264_CHROMA_MC(put_ , op_put)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1449 H264_CHROMA_MC(avg_ , op_avg)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1450 #undef op_avg
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1451 #undef op_put
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1452
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1453 static inline void copy_block4(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1454 {
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1455 int i;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1456 for(i=0; i<h; i++)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1457 {
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1458 ST32(dst , LD32(src ));
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1459 dst+=dstStride;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1460 src+=srcStride;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1461 }
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1462 }
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1463
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1464 static inline void copy_block8(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1465 {
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1466 int i;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1467 for(i=0; i<h; i++)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1468 {
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1469 ST32(dst , LD32(src ));
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1470 ST32(dst+4 , LD32(src+4 ));
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1471 dst+=dstStride;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1472 src+=srcStride;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1473 }
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1474 }
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1475
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1476 static inline void copy_block16(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1477 {
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1478 int i;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1479 for(i=0; i<h; i++)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1480 {
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1481 ST32(dst , LD32(src ));
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1482 ST32(dst+4 , LD32(src+4 ));
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1483 ST32(dst+8 , LD32(src+8 ));
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1484 ST32(dst+12, LD32(src+12));
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1485 dst+=dstStride;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1486 src+=srcStride;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1487 }
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1488 }
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1489
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1490 static inline void copy_block17(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h)
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1491 {
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1492 int i;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1493 for(i=0; i<h; i++)
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1494 {
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1495 ST32(dst , LD32(src ));
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1496 ST32(dst+4 , LD32(src+4 ));
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1497 ST32(dst+8 , LD32(src+8 ));
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1498 ST32(dst+12, LD32(src+12));
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1499 dst[16]= src[16];
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1500 dst+=dstStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1501 src+=srcStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1502 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1503 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1504
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1505 static inline void copy_block9(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h)
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1506 {
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1507 int i;
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1508 for(i=0; i<h; i++)
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1509 {
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1510 ST32(dst , LD32(src ));
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1511 ST32(dst+4 , LD32(src+4 ));
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1512 dst[8]= src[8];
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1513 dst+=dstStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1514 src+=srcStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1515 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1516 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1517
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1518
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1519 #define QPEL_MC(r, OPNAME, RND, OP) \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1520 static void OPNAME ## mpeg4_qpel8_h_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1521 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1522 int i;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1523 for(i=0; i<h; i++)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1524 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1525 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
1526 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
1527 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
1528 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
1529 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
1530 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
1531 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
1532 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
1533 dst+=dstStride;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1534 src+=srcStride;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1535 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1536 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1537 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1538 static void OPNAME ## mpeg4_qpel8_v_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1539 const int w=8;\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1540 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1541 int i;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1542 for(i=0; i<w; i++)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1543 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1544 const int src0= src[0*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1545 const int src1= src[1*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1546 const int src2= src[2*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1547 const int src3= src[3*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1548 const int src4= src[4*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1549 const int src5= src[5*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1550 const int src6= src[6*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1551 const int src7= src[7*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1552 const int src8= src[8*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1553 OP(dst[0*dstStride], (src0+src1)*20 - (src0+src2)*6 + (src1+src3)*3 - (src2+src4));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1554 OP(dst[1*dstStride], (src1+src2)*20 - (src0+src3)*6 + (src0+src4)*3 - (src1+src5));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1555 OP(dst[2*dstStride], (src2+src3)*20 - (src1+src4)*6 + (src0+src5)*3 - (src0+src6));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1556 OP(dst[3*dstStride], (src3+src4)*20 - (src2+src5)*6 + (src1+src6)*3 - (src0+src7));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1557 OP(dst[4*dstStride], (src4+src5)*20 - (src3+src6)*6 + (src2+src7)*3 - (src1+src8));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1558 OP(dst[5*dstStride], (src5+src6)*20 - (src4+src7)*6 + (src3+src8)*3 - (src2+src8));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1559 OP(dst[6*dstStride], (src6+src7)*20 - (src5+src8)*6 + (src4+src8)*3 - (src3+src7));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1560 OP(dst[7*dstStride], (src7+src8)*20 - (src6+src8)*6 + (src5+src7)*3 - (src4+src6));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1561 dst++;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1562 src++;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1563 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1564 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1565 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1566 static void OPNAME ## mpeg4_qpel16_h_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1567 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1568 int i;\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1569 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1570 for(i=0; i<h; i++)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1571 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1572 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
1573 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
1574 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
1575 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
1576 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
1577 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
1578 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
1579 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
1580 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
1581 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
1582 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
1583 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
1584 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
1585 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
1586 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
1587 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
1588 dst+=dstStride;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1589 src+=srcStride;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1590 }\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1591 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1592 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1593 static void OPNAME ## mpeg4_qpel16_v_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1594 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1595 int i;\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1596 const int w=16;\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1597 for(i=0; i<w; i++)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1598 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1599 const int src0= src[0*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1600 const int src1= src[1*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1601 const int src2= src[2*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1602 const int src3= src[3*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1603 const int src4= src[4*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1604 const int src5= src[5*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1605 const int src6= src[6*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1606 const int src7= src[7*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1607 const int src8= src[8*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1608 const int src9= src[9*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1609 const int src10= src[10*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1610 const int src11= src[11*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1611 const int src12= src[12*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1612 const int src13= src[13*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1613 const int src14= src[14*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1614 const int src15= src[15*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1615 const int src16= src[16*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1616 OP(dst[ 0*dstStride], (src0 +src1 )*20 - (src0 +src2 )*6 + (src1 +src3 )*3 - (src2 +src4 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1617 OP(dst[ 1*dstStride], (src1 +src2 )*20 - (src0 +src3 )*6 + (src0 +src4 )*3 - (src1 +src5 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1618 OP(dst[ 2*dstStride], (src2 +src3 )*20 - (src1 +src4 )*6 + (src0 +src5 )*3 - (src0 +src6 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1619 OP(dst[ 3*dstStride], (src3 +src4 )*20 - (src2 +src5 )*6 + (src1 +src6 )*3 - (src0 +src7 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1620 OP(dst[ 4*dstStride], (src4 +src5 )*20 - (src3 +src6 )*6 + (src2 +src7 )*3 - (src1 +src8 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1621 OP(dst[ 5*dstStride], (src5 +src6 )*20 - (src4 +src7 )*6 + (src3 +src8 )*3 - (src2 +src9 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1622 OP(dst[ 6*dstStride], (src6 +src7 )*20 - (src5 +src8 )*6 + (src4 +src9 )*3 - (src3 +src10));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1623 OP(dst[ 7*dstStride], (src7 +src8 )*20 - (src6 +src9 )*6 + (src5 +src10)*3 - (src4 +src11));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1624 OP(dst[ 8*dstStride], (src8 +src9 )*20 - (src7 +src10)*6 + (src6 +src11)*3 - (src5 +src12));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1625 OP(dst[ 9*dstStride], (src9 +src10)*20 - (src8 +src11)*6 + (src7 +src12)*3 - (src6 +src13));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1626 OP(dst[10*dstStride], (src10+src11)*20 - (src9 +src12)*6 + (src8 +src13)*3 - (src7 +src14));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1627 OP(dst[11*dstStride], (src11+src12)*20 - (src10+src13)*6 + (src9 +src14)*3 - (src8 +src15));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1628 OP(dst[12*dstStride], (src12+src13)*20 - (src11+src14)*6 + (src10+src15)*3 - (src9 +src16));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1629 OP(dst[13*dstStride], (src13+src14)*20 - (src12+src15)*6 + (src11+src16)*3 - (src10+src16));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1630 OP(dst[14*dstStride], (src14+src15)*20 - (src13+src16)*6 + (src12+src16)*3 - (src11+src15));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1631 OP(dst[15*dstStride], (src15+src16)*20 - (src14+src16)*6 + (src13+src15)*3 - (src12+src14));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1632 dst++;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1633 src++;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1634 }\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1635 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1636 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1637 static void OPNAME ## qpel8_mc00_c (uint8_t *dst, uint8_t *src, int stride){\
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1638 OPNAME ## pixels8_c(dst, src, stride, 8);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1639 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1640 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1641 static void OPNAME ## qpel8_mc10_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1642 uint8_t half[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1643 put ## RND ## mpeg4_qpel8_h_lowpass(half, src, 8, stride, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1644 OPNAME ## pixels8_l2(dst, src, half, stride, stride, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1645 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1646 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1647 static void OPNAME ## qpel8_mc20_c(uint8_t *dst, uint8_t *src, int stride){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1648 OPNAME ## mpeg4_qpel8_h_lowpass(dst, src, stride, stride, 8);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1649 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1650 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1651 static void OPNAME ## qpel8_mc30_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1652 uint8_t half[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1653 put ## RND ## mpeg4_qpel8_h_lowpass(half, src, 8, stride, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1654 OPNAME ## pixels8_l2(dst, src+1, half, stride, stride, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1655 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1656 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1657 static void OPNAME ## qpel8_mc01_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1658 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1659 uint8_t half[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1660 copy_block9(full, src, 16, stride, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1661 put ## RND ## mpeg4_qpel8_v_lowpass(half, full, 8, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1662 OPNAME ## pixels8_l2(dst, full, half, stride, 16, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1663 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1664 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1665 static void OPNAME ## qpel8_mc02_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1666 uint8_t full[16*9];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1667 copy_block9(full, src, 16, stride, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1668 OPNAME ## mpeg4_qpel8_v_lowpass(dst, full, stride, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1669 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1670 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1671 static void OPNAME ## qpel8_mc03_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1672 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1673 uint8_t half[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1674 copy_block9(full, src, 16, stride, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1675 put ## RND ## mpeg4_qpel8_v_lowpass(half, full, 8, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1676 OPNAME ## pixels8_l2(dst, full+16, half, stride, 16, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1677 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1678 void ff_ ## OPNAME ## qpel8_mc11_old_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1679 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1680 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1681 uint8_t halfV[64];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1682 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1683 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1684 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1685 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full, 8, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1686 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1687 OPNAME ## pixels8_l4(dst, full, halfH, halfV, halfHV, stride, 16, 8, 8, 8, 8);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1688 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1689 static void OPNAME ## qpel8_mc11_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1690 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1691 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1692 uint8_t halfHV[64];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1693 copy_block9(full, src, 16, stride, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1694 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1695 put ## RND ## pixels8_l2(halfH, halfH, full, 8, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1696 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1697 OPNAME ## pixels8_l2(dst, halfH, halfHV, stride, 8, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1698 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1699 void ff_ ## OPNAME ## qpel8_mc31_old_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1700 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1701 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1702 uint8_t halfV[64];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1703 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1704 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1705 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1706 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full+1, 8, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1707 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1708 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
1709 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1710 static void OPNAME ## qpel8_mc31_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1711 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1712 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1713 uint8_t halfHV[64];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1714 copy_block9(full, src, 16, stride, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1715 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1716 put ## RND ## pixels8_l2(halfH, halfH, full+1, 8, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1717 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1718 OPNAME ## pixels8_l2(dst, halfH, halfHV, stride, 8, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1719 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1720 void ff_ ## OPNAME ## qpel8_mc13_old_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1721 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1722 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1723 uint8_t halfV[64];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1724 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1725 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1726 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1727 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full, 8, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1728 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1729 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
1730 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1731 static void OPNAME ## qpel8_mc13_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1732 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1733 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1734 uint8_t halfHV[64];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1735 copy_block9(full, src, 16, stride, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1736 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1737 put ## RND ## pixels8_l2(halfH, halfH, full, 8, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1738 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1739 OPNAME ## pixels8_l2(dst, halfH+8, halfHV, stride, 8, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1740 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1741 void ff_ ## OPNAME ## qpel8_mc33_old_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1742 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1743 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1744 uint8_t halfV[64];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1745 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1746 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1747 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full , 8, 16, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1748 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full+1, 8, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1749 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1750 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
1751 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1752 static void OPNAME ## qpel8_mc33_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1753 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1754 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1755 uint8_t halfHV[64];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1756 copy_block9(full, src, 16, stride, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1757 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1758 put ## RND ## pixels8_l2(halfH, halfH, full+1, 8, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1759 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1760 OPNAME ## pixels8_l2(dst, halfH+8, halfHV, stride, 8, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1761 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1762 static void OPNAME ## qpel8_mc21_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1763 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1764 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1765 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, src, 8, stride, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1766 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1767 OPNAME ## pixels8_l2(dst, halfH, halfHV, stride, 8, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1768 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1769 static void OPNAME ## qpel8_mc23_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1770 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1771 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1772 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, src, 8, stride, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1773 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1774 OPNAME ## pixels8_l2(dst, halfH+8, halfHV, stride, 8, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1775 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1776 void ff_ ## OPNAME ## qpel8_mc12_old_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1777 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1778 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1779 uint8_t halfV[64];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1780 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1781 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1782 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1783 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full, 8, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1784 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1785 OPNAME ## pixels8_l2(dst, halfV, halfHV, stride, 8, 8, 8);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1786 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1787 static void OPNAME ## qpel8_mc12_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1788 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1789 uint8_t halfH[72];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1790 copy_block9(full, src, 16, stride, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1791 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1792 put ## RND ## pixels8_l2(halfH, halfH, full, 8, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1793 OPNAME ## mpeg4_qpel8_v_lowpass(dst, halfH, stride, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1794 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1795 void ff_ ## OPNAME ## qpel8_mc32_old_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1796 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1797 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1798 uint8_t halfV[64];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1799 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1800 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1801 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1802 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full+1, 8, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1803 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1804 OPNAME ## pixels8_l2(dst, halfV, halfHV, stride, 8, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1805 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1806 static void OPNAME ## qpel8_mc32_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1807 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1808 uint8_t halfH[72];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1809 copy_block9(full, src, 16, stride, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1810 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1811 put ## RND ## pixels8_l2(halfH, halfH, full+1, 8, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1812 OPNAME ## mpeg4_qpel8_v_lowpass(dst, halfH, stride, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1813 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1814 static void OPNAME ## qpel8_mc22_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1815 uint8_t halfH[72];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1816 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, src, 8, stride, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1817 OPNAME ## mpeg4_qpel8_v_lowpass(dst, halfH, stride, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1818 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1819 static void OPNAME ## qpel16_mc00_c (uint8_t *dst, uint8_t *src, int stride){\
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1820 OPNAME ## pixels16_c(dst, src, stride, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1821 }\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1822 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1823 static void OPNAME ## qpel16_mc10_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1824 uint8_t half[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1825 put ## RND ## mpeg4_qpel16_h_lowpass(half, src, 16, stride, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1826 OPNAME ## pixels16_l2(dst, src, half, stride, stride, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1827 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1828 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1829 static void OPNAME ## qpel16_mc20_c(uint8_t *dst, uint8_t *src, int stride){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1830 OPNAME ## mpeg4_qpel16_h_lowpass(dst, src, stride, stride, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1831 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1832 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1833 static void OPNAME ## qpel16_mc30_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1834 uint8_t half[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1835 put ## RND ## mpeg4_qpel16_h_lowpass(half, src, 16, stride, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1836 OPNAME ## pixels16_l2(dst, src+1, half, stride, stride, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1837 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1838 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1839 static void OPNAME ## qpel16_mc01_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1840 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1841 uint8_t half[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1842 copy_block17(full, src, 24, stride, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1843 put ## RND ## mpeg4_qpel16_v_lowpass(half, full, 16, 24);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1844 OPNAME ## pixels16_l2(dst, full, half, stride, 24, 16, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1845 }\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1846 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1847 static void OPNAME ## qpel16_mc02_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1848 uint8_t full[24*17];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1849 copy_block17(full, src, 24, stride, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1850 OPNAME ## mpeg4_qpel16_v_lowpass(dst, full, stride, 24);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1851 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1852 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1853 static void OPNAME ## qpel16_mc03_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1854 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1855 uint8_t half[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1856 copy_block17(full, src, 24, stride, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1857 put ## RND ## mpeg4_qpel16_v_lowpass(half, full, 16, 24);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1858 OPNAME ## pixels16_l2(dst, full+24, half, stride, 24, 16, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1859 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1860 void ff_ ## OPNAME ## qpel16_mc11_old_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1861 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1862 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1863 uint8_t halfV[256];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1864 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1865 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1866 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1867 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full, 16, 24);\
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1868 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1869 OPNAME ## pixels16_l4(dst, full, halfH, halfV, halfHV, stride, 24, 16, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1870 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1871 static void OPNAME ## qpel16_mc11_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1872 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1873 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1874 uint8_t halfHV[256];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1875 copy_block17(full, src, 24, stride, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1876 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1877 put ## RND ## pixels16_l2(halfH, halfH, full, 16, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1878 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1879 OPNAME ## pixels16_l2(dst, halfH, halfHV, stride, 16, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1880 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1881 void ff_ ## OPNAME ## qpel16_mc31_old_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1882 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1883 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1884 uint8_t halfV[256];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1885 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1886 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1887 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1888 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full+1, 16, 24);\
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1889 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1890 OPNAME ## pixels16_l4(dst, full+1, halfH, halfV, halfHV, stride, 24, 16, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1891 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1892 static void OPNAME ## qpel16_mc31_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1893 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1894 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1895 uint8_t halfHV[256];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1896 copy_block17(full, src, 24, stride, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1897 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1898 put ## RND ## pixels16_l2(halfH, halfH, full+1, 16, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1899 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1900 OPNAME ## pixels16_l2(dst, halfH, halfHV, stride, 16, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1901 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1902 void ff_ ## OPNAME ## qpel16_mc13_old_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1903 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1904 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1905 uint8_t halfV[256];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1906 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1907 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1908 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1909 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full, 16, 24);\
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1910 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1911 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
1912 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1913 static void OPNAME ## qpel16_mc13_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1914 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1915 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1916 uint8_t halfHV[256];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1917 copy_block17(full, src, 24, stride, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1918 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1919 put ## RND ## pixels16_l2(halfH, halfH, full, 16, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1920 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1921 OPNAME ## pixels16_l2(dst, halfH+16, halfHV, stride, 16, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1922 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1923 void ff_ ## OPNAME ## qpel16_mc33_old_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1924 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1925 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1926 uint8_t halfV[256];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1927 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1928 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1929 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full , 16, 24, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1930 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full+1, 16, 24);\
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1931 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1932 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
1933 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1934 static void OPNAME ## qpel16_mc33_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1935 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1936 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1937 uint8_t halfHV[256];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1938 copy_block17(full, src, 24, stride, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1939 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1940 put ## RND ## pixels16_l2(halfH, halfH, full+1, 16, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1941 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1942 OPNAME ## pixels16_l2(dst, halfH+16, halfHV, stride, 16, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1943 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1944 static void OPNAME ## qpel16_mc21_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1945 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1946 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1947 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, src, 16, stride, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1948 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1949 OPNAME ## pixels16_l2(dst, halfH, halfHV, stride, 16, 16, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1950 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1951 static void OPNAME ## qpel16_mc23_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1952 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1953 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1954 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, src, 16, stride, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1955 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1956 OPNAME ## pixels16_l2(dst, halfH+16, halfHV, stride, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1957 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1958 void ff_ ## OPNAME ## qpel16_mc12_old_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1959 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1960 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1961 uint8_t halfV[256];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1962 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1963 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1964 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1965 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full, 16, 24);\
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1966 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1967 OPNAME ## pixels16_l2(dst, halfV, halfHV, stride, 16, 16, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1968 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1969 static void OPNAME ## qpel16_mc12_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1970 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1971 uint8_t halfH[272];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1972 copy_block17(full, src, 24, stride, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1973 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1974 put ## RND ## pixels16_l2(halfH, halfH, full, 16, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1975 OPNAME ## mpeg4_qpel16_v_lowpass(dst, halfH, stride, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1976 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1977 void ff_ ## OPNAME ## qpel16_mc32_old_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1978 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1979 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1980 uint8_t halfV[256];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1981 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1982 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1983 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1984 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full+1, 16, 24);\
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1985 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1986 OPNAME ## pixels16_l2(dst, halfV, halfHV, stride, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1987 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1988 static void OPNAME ## qpel16_mc32_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1989 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1990 uint8_t halfH[272];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1991 copy_block17(full, src, 24, stride, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1992 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1993 put ## RND ## pixels16_l2(halfH, halfH, full+1, 16, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1994 OPNAME ## mpeg4_qpel16_v_lowpass(dst, halfH, stride, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1995 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1996 static void OPNAME ## qpel16_mc22_c(uint8_t *dst, uint8_t *src, int stride){\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1997 uint8_t halfH[272];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1998 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, src, 16, stride, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1999 OPNAME ## mpeg4_qpel16_v_lowpass(dst, halfH, stride, 16);\
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
2000 }
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
2001
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2002 #define op_avg(a, b) a = (((a)+cm[((b) + 16)>>5]+1)>>1)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2003 #define op_avg_no_rnd(a, b) a = (((a)+cm[((b) + 15)>>5])>>1)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2004 #define op_put(a, b) a = cm[((b) + 16)>>5]
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2005 #define op_put_no_rnd(a, b) a = cm[((b) + 15)>>5]
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2006
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2007 QPEL_MC(0, put_ , _ , op_put)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2008 QPEL_MC(1, put_no_rnd_, _no_rnd_, op_put_no_rnd)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2009 QPEL_MC(0, avg_ , _ , op_avg)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2010 //QPEL_MC(1, avg_no_rnd , _ , op_avg)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2011 #undef op_avg
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2012 #undef op_avg_no_rnd
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2013 #undef op_put
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2014 #undef op_put_no_rnd
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
2015
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2016 #if 1
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2017 #define H264_LOWPASS(OPNAME, OP, OP2) \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2018 static void OPNAME ## h264_qpel4_h_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2019 const int h=4;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2020 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2021 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2022 for(i=0; i<h; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2023 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2024 OP(dst[0], (src[0]+src[1])*20 - (src[-1]+src[2])*5 + (src[-2]+src[3]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2025 OP(dst[1], (src[1]+src[2])*20 - (src[0 ]+src[3])*5 + (src[-1]+src[4]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2026 OP(dst[2], (src[2]+src[3])*20 - (src[1 ]+src[4])*5 + (src[0 ]+src[5]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2027 OP(dst[3], (src[3]+src[4])*20 - (src[2 ]+src[5])*5 + (src[1 ]+src[6]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2028 dst+=dstStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2029 src+=srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2030 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2031 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2032 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2033 static void OPNAME ## h264_qpel4_v_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2034 const int w=4;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2035 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2036 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2037 for(i=0; i<w; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2038 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2039 const int srcB= src[-2*srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2040 const int srcA= src[-1*srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2041 const int src0= src[0 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2042 const int src1= src[1 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2043 const int src2= src[2 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2044 const int src3= src[3 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2045 const int src4= src[4 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2046 const int src5= src[5 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2047 const int src6= src[6 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2048 OP(dst[0*dstStride], (src0+src1)*20 - (srcA+src2)*5 + (srcB+src3));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2049 OP(dst[1*dstStride], (src1+src2)*20 - (src0+src3)*5 + (srcA+src4));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2050 OP(dst[2*dstStride], (src2+src3)*20 - (src1+src4)*5 + (src0+src5));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2051 OP(dst[3*dstStride], (src3+src4)*20 - (src2+src5)*5 + (src1+src6));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2052 dst++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2053 src++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2054 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2055 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2056 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2057 static void OPNAME ## h264_qpel4_hv_lowpass(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2058 const int h=4;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2059 const int w=4;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2060 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2061 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2062 src -= 2*srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2063 for(i=0; i<h+5; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2064 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2065 tmp[0]= (src[0]+src[1])*20 - (src[-1]+src[2])*5 + (src[-2]+src[3]);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2066 tmp[1]= (src[1]+src[2])*20 - (src[0 ]+src[3])*5 + (src[-1]+src[4]);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2067 tmp[2]= (src[2]+src[3])*20 - (src[1 ]+src[4])*5 + (src[0 ]+src[5]);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2068 tmp[3]= (src[3]+src[4])*20 - (src[2 ]+src[5])*5 + (src[1 ]+src[6]);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2069 tmp+=tmpStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2070 src+=srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2071 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2072 tmp -= tmpStride*(h+5-2);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2073 for(i=0; i<w; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2074 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2075 const int tmpB= tmp[-2*tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2076 const int tmpA= tmp[-1*tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2077 const int tmp0= tmp[0 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2078 const int tmp1= tmp[1 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2079 const int tmp2= tmp[2 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2080 const int tmp3= tmp[3 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2081 const int tmp4= tmp[4 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2082 const int tmp5= tmp[5 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2083 const int tmp6= tmp[6 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2084 OP2(dst[0*dstStride], (tmp0+tmp1)*20 - (tmpA+tmp2)*5 + (tmpB+tmp3));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2085 OP2(dst[1*dstStride], (tmp1+tmp2)*20 - (tmp0+tmp3)*5 + (tmpA+tmp4));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2086 OP2(dst[2*dstStride], (tmp2+tmp3)*20 - (tmp1+tmp4)*5 + (tmp0+tmp5));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2087 OP2(dst[3*dstStride], (tmp3+tmp4)*20 - (tmp2+tmp5)*5 + (tmp1+tmp6));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2088 dst++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2089 tmp++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2090 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2091 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2092 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2093 static void OPNAME ## h264_qpel8_h_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2094 const int h=8;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2095 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2096 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2097 for(i=0; i<h; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2098 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2099 OP(dst[0], (src[0]+src[1])*20 - (src[-1]+src[2])*5 + (src[-2]+src[3 ]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2100 OP(dst[1], (src[1]+src[2])*20 - (src[0 ]+src[3])*5 + (src[-1]+src[4 ]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2101 OP(dst[2], (src[2]+src[3])*20 - (src[1 ]+src[4])*5 + (src[0 ]+src[5 ]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2102 OP(dst[3], (src[3]+src[4])*20 - (src[2 ]+src[5])*5 + (src[1 ]+src[6 ]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2103 OP(dst[4], (src[4]+src[5])*20 - (src[3 ]+src[6])*5 + (src[2 ]+src[7 ]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2104 OP(dst[5], (src[5]+src[6])*20 - (src[4 ]+src[7])*5 + (src[3 ]+src[8 ]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2105 OP(dst[6], (src[6]+src[7])*20 - (src[5 ]+src[8])*5 + (src[4 ]+src[9 ]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2106 OP(dst[7], (src[7]+src[8])*20 - (src[6 ]+src[9])*5 + (src[5 ]+src[10]));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2107 dst+=dstStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2108 src+=srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2109 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2110 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2111 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2112 static void OPNAME ## h264_qpel8_v_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2113 const int w=8;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2114 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2115 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2116 for(i=0; i<w; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2117 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2118 const int srcB= src[-2*srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2119 const int srcA= src[-1*srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2120 const int src0= src[0 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2121 const int src1= src[1 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2122 const int src2= src[2 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2123 const int src3= src[3 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2124 const int src4= src[4 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2125 const int src5= src[5 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2126 const int src6= src[6 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2127 const int src7= src[7 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2128 const int src8= src[8 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2129 const int src9= src[9 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2130 const int src10=src[10*srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2131 OP(dst[0*dstStride], (src0+src1)*20 - (srcA+src2)*5 + (srcB+src3));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2132 OP(dst[1*dstStride], (src1+src2)*20 - (src0+src3)*5 + (srcA+src4));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2133 OP(dst[2*dstStride], (src2+src3)*20 - (src1+src4)*5 + (src0+src5));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2134 OP(dst[3*dstStride], (src3+src4)*20 - (src2+src5)*5 + (src1+src6));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2135 OP(dst[4*dstStride], (src4+src5)*20 - (src3+src6)*5 + (src2+src7));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2136 OP(dst[5*dstStride], (src5+src6)*20 - (src4+src7)*5 + (src3+src8));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2137 OP(dst[6*dstStride], (src6+src7)*20 - (src5+src8)*5 + (src4+src9));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2138 OP(dst[7*dstStride], (src7+src8)*20 - (src6+src9)*5 + (src5+src10));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2139 dst++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2140 src++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2141 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2142 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2143 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2144 static void OPNAME ## h264_qpel8_hv_lowpass(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2145 const int h=8;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2146 const int w=8;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2147 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2148 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2149 src -= 2*srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2150 for(i=0; i<h+5; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2151 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2152 tmp[0]= (src[0]+src[1])*20 - (src[-1]+src[2])*5 + (src[-2]+src[3 ]);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2153 tmp[1]= (src[1]+src[2])*20 - (src[0 ]+src[3])*5 + (src[-1]+src[4 ]);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2154 tmp[2]= (src[2]+src[3])*20 - (src[1 ]+src[4])*5 + (src[0 ]+src[5 ]);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2155 tmp[3]= (src[3]+src[4])*20 - (src[2 ]+src[5])*5 + (src[1 ]+src[6 ]);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2156 tmp[4]= (src[4]+src[5])*20 - (src[3 ]+src[6])*5 + (src[2 ]+src[7 ]);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2157 tmp[5]= (src[5]+src[6])*20 - (src[4 ]+src[7])*5 + (src[3 ]+src[8 ]);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2158 tmp[6]= (src[6]+src[7])*20 - (src[5 ]+src[8])*5 + (src[4 ]+src[9 ]);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2159 tmp[7]= (src[7]+src[8])*20 - (src[6 ]+src[9])*5 + (src[5 ]+src[10]);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2160 tmp+=tmpStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2161 src+=srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2162 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2163 tmp -= tmpStride*(h+5-2);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2164 for(i=0; i<w; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2165 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2166 const int tmpB= tmp[-2*tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2167 const int tmpA= tmp[-1*tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2168 const int tmp0= tmp[0 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2169 const int tmp1= tmp[1 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2170 const int tmp2= tmp[2 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2171 const int tmp3= tmp[3 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2172 const int tmp4= tmp[4 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2173 const int tmp5= tmp[5 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2174 const int tmp6= tmp[6 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2175 const int tmp7= tmp[7 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2176 const int tmp8= tmp[8 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2177 const int tmp9= tmp[9 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2178 const int tmp10=tmp[10*tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2179 OP2(dst[0*dstStride], (tmp0+tmp1)*20 - (tmpA+tmp2)*5 + (tmpB+tmp3));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2180 OP2(dst[1*dstStride], (tmp1+tmp2)*20 - (tmp0+tmp3)*5 + (tmpA+tmp4));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2181 OP2(dst[2*dstStride], (tmp2+tmp3)*20 - (tmp1+tmp4)*5 + (tmp0+tmp5));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2182 OP2(dst[3*dstStride], (tmp3+tmp4)*20 - (tmp2+tmp5)*5 + (tmp1+tmp6));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2183 OP2(dst[4*dstStride], (tmp4+tmp5)*20 - (tmp3+tmp6)*5 + (tmp2+tmp7));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2184 OP2(dst[5*dstStride], (tmp5+tmp6)*20 - (tmp4+tmp7)*5 + (tmp3+tmp8));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2185 OP2(dst[6*dstStride], (tmp6+tmp7)*20 - (tmp5+tmp8)*5 + (tmp4+tmp9));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2186 OP2(dst[7*dstStride], (tmp7+tmp8)*20 - (tmp6+tmp9)*5 + (tmp5+tmp10));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2187 dst++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2188 tmp++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2189 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2190 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2191 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2192 static void OPNAME ## h264_qpel16_v_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2193 OPNAME ## h264_qpel8_v_lowpass(dst , src , dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2194 OPNAME ## h264_qpel8_v_lowpass(dst+8, src+8, dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2195 src += 8*srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2196 dst += 8*dstStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2197 OPNAME ## h264_qpel8_v_lowpass(dst , src , dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2198 OPNAME ## h264_qpel8_v_lowpass(dst+8, src+8, dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2199 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2200 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2201 static void OPNAME ## h264_qpel16_h_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2202 OPNAME ## h264_qpel8_h_lowpass(dst , src , dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2203 OPNAME ## h264_qpel8_h_lowpass(dst+8, src+8, dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2204 src += 8*srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2205 dst += 8*dstStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2206 OPNAME ## h264_qpel8_h_lowpass(dst , src , dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2207 OPNAME ## h264_qpel8_h_lowpass(dst+8, src+8, dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2208 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2209 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2210 static void OPNAME ## h264_qpel16_hv_lowpass(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2211 OPNAME ## h264_qpel8_hv_lowpass(dst , tmp , src , dstStride, tmpStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2212 OPNAME ## h264_qpel8_hv_lowpass(dst+8, tmp+8, src+8, dstStride, tmpStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2213 src += 8*srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2214 dst += 8*dstStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2215 OPNAME ## h264_qpel8_hv_lowpass(dst , tmp , src , dstStride, tmpStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2216 OPNAME ## h264_qpel8_hv_lowpass(dst+8, tmp+8, src+8, dstStride, tmpStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2217 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2218
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2219 #define H264_MC(OPNAME, SIZE) \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2220 static void OPNAME ## h264_qpel ## SIZE ## _mc00_c (uint8_t *dst, uint8_t *src, int stride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2221 OPNAME ## pixels ## SIZE ## _c(dst, src, stride, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2222 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2223 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2224 static void OPNAME ## h264_qpel ## SIZE ## _mc10_c(uint8_t *dst, uint8_t *src, int stride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2225 uint8_t half[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2226 put_h264_qpel ## SIZE ## _h_lowpass(half, src, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2227 OPNAME ## pixels ## SIZE ## _l2(dst, src, half, stride, stride, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2228 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2229 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2230 static void OPNAME ## h264_qpel ## SIZE ## _mc20_c(uint8_t *dst, uint8_t *src, int stride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2231 OPNAME ## h264_qpel ## SIZE ## _h_lowpass(dst, src, stride, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2232 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2233 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2234 static void OPNAME ## h264_qpel ## SIZE ## _mc30_c(uint8_t *dst, uint8_t *src, int stride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2235 uint8_t half[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2236 put_h264_qpel ## SIZE ## _h_lowpass(half, src, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2237 OPNAME ## pixels ## SIZE ## _l2(dst, src+1, half, stride, stride, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2238 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2239 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2240 static void OPNAME ## h264_qpel ## SIZE ## _mc01_c(uint8_t *dst, uint8_t *src, int stride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2241 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2242 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2243 uint8_t half[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2244 copy_block ## SIZE (full, src - stride*2, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2245 put_h264_qpel ## SIZE ## _v_lowpass(half, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2246 OPNAME ## pixels ## SIZE ## _l2(dst, full_mid, half, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2247 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2248 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2249 static void OPNAME ## h264_qpel ## SIZE ## _mc02_c(uint8_t *dst, uint8_t *src, int stride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2250 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2251 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2252 copy_block ## SIZE (full, src - stride*2, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2253 OPNAME ## h264_qpel ## SIZE ## _v_lowpass(dst, full_mid, stride, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2254 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2255 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2256 static void OPNAME ## h264_qpel ## SIZE ## _mc03_c(uint8_t *dst, uint8_t *src, int stride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2257 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2258 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2259 uint8_t half[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2260 copy_block ## SIZE (full, src - stride*2, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2261 put_h264_qpel ## SIZE ## _v_lowpass(half, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2262 OPNAME ## pixels ## SIZE ## _l2(dst, full_mid+SIZE, half, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2263 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2264 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2265 static void OPNAME ## h264_qpel ## SIZE ## _mc11_c(uint8_t *dst, uint8_t *src, int stride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2266 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2267 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2268 uint8_t halfH[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2269 uint8_t halfV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2270 put_h264_qpel ## SIZE ## _h_lowpass(halfH, src, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2271 copy_block ## SIZE (full, src - stride*2, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2272 put_h264_qpel ## SIZE ## _v_lowpass(halfV, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2273 OPNAME ## pixels ## SIZE ## _l2(dst, halfH, halfV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2274 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2275 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2276 static void OPNAME ## h264_qpel ## SIZE ## _mc31_c(uint8_t *dst, uint8_t *src, int stride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2277 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2278 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2279 uint8_t halfH[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2280 uint8_t halfV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2281 put_h264_qpel ## SIZE ## _h_lowpass(halfH, src, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2282 copy_block ## SIZE (full, src - stride*2 + 1, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2283 put_h264_qpel ## SIZE ## _v_lowpass(halfV, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2284 OPNAME ## pixels ## SIZE ## _l2(dst, halfH, halfV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2285 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2286 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2287 static void OPNAME ## h264_qpel ## SIZE ## _mc13_c(uint8_t *dst, uint8_t *src, int stride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2288 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2289 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2290 uint8_t halfH[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2291 uint8_t halfV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2292 put_h264_qpel ## SIZE ## _h_lowpass(halfH, src + stride, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2293 copy_block ## SIZE (full, src - stride*2, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2294 put_h264_qpel ## SIZE ## _v_lowpass(halfV, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2295 OPNAME ## pixels ## SIZE ## _l2(dst, halfH, halfV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2296 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2297 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2298 static void OPNAME ## h264_qpel ## SIZE ## _mc33_c(uint8_t *dst, uint8_t *src, int stride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2299 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2300 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2301 uint8_t halfH[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2302 uint8_t halfV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2303 put_h264_qpel ## SIZE ## _h_lowpass(halfH, src + stride, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2304 copy_block ## SIZE (full, src - stride*2 + 1, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2305 put_h264_qpel ## SIZE ## _v_lowpass(halfV, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2306 OPNAME ## pixels ## SIZE ## _l2(dst, halfH, halfV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2307 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2308 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2309 static void OPNAME ## h264_qpel ## SIZE ## _mc22_c(uint8_t *dst, uint8_t *src, int stride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2310 int16_t tmp[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2311 OPNAME ## h264_qpel ## SIZE ## _hv_lowpass(dst, tmp, src, stride, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2312 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2313 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2314 static void OPNAME ## h264_qpel ## SIZE ## _mc21_c(uint8_t *dst, uint8_t *src, int stride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2315 int16_t tmp[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2316 uint8_t halfH[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2317 uint8_t halfHV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2318 put_h264_qpel ## SIZE ## _h_lowpass(halfH, src, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2319 put_h264_qpel ## SIZE ## _hv_lowpass(halfHV, tmp, src, SIZE, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2320 OPNAME ## pixels ## SIZE ## _l2(dst, halfH, halfHV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2321 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2322 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2323 static void OPNAME ## h264_qpel ## SIZE ## _mc23_c(uint8_t *dst, uint8_t *src, int stride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2324 int16_t tmp[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2325 uint8_t halfH[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2326 uint8_t halfHV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2327 put_h264_qpel ## SIZE ## _h_lowpass(halfH, src + stride, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2328 put_h264_qpel ## SIZE ## _hv_lowpass(halfHV, tmp, src, SIZE, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2329 OPNAME ## pixels ## SIZE ## _l2(dst, halfH, halfHV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2330 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2331 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2332 static void OPNAME ## h264_qpel ## SIZE ## _mc12_c(uint8_t *dst, uint8_t *src, int stride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2333 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2334 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2335 int16_t tmp[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2336 uint8_t halfV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2337 uint8_t halfHV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2338 copy_block ## SIZE (full, src - stride*2, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2339 put_h264_qpel ## SIZE ## _v_lowpass(halfV, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2340 put_h264_qpel ## SIZE ## _hv_lowpass(halfHV, tmp, src, SIZE, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2341 OPNAME ## pixels ## SIZE ## _l2(dst, halfV, halfHV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2342 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2343 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2344 static void OPNAME ## h264_qpel ## SIZE ## _mc32_c(uint8_t *dst, uint8_t *src, int stride){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2345 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2346 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2347 int16_t tmp[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2348 uint8_t halfV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2349 uint8_t halfHV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2350 copy_block ## SIZE (full, src - stride*2 + 1, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2351 put_h264_qpel ## SIZE ## _v_lowpass(halfV, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2352 put_h264_qpel ## SIZE ## _hv_lowpass(halfHV, tmp, src, SIZE, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2353 OPNAME ## pixels ## SIZE ## _l2(dst, halfV, halfHV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2354 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2355
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2356 #define op_avg(a, b) a = (((a)+cm[((b) + 16)>>5]+1)>>1)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2357 //#define op_avg2(a, b) a = (((a)*w1+cm[((b) + 16)>>5]*w2 + o + 64)>>7)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2358 #define op_put(a, b) a = cm[((b) + 16)>>5]
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2359 #define op2_avg(a, b) a = (((a)+cm[((b) + 512)>>10]+1)>>1)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2360 #define op2_put(a, b) a = cm[((b) + 512)>>10]
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2361
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2362 H264_LOWPASS(put_ , op_put, op2_put)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2363 H264_LOWPASS(avg_ , op_avg, op2_avg)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2364 H264_MC(put_, 4)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2365 H264_MC(put_, 8)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2366 H264_MC(put_, 16)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2367 H264_MC(avg_, 4)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2368 H264_MC(avg_, 8)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2369 H264_MC(avg_, 16)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2370
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2371 #undef op_avg
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2372 #undef op_put
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2373 #undef op2_avg
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2374 #undef op2_put
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2375 #endif
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2376
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2377 static void wmv2_mspel8_h_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2378 uint8_t *cm = cropTbl + MAX_NEG_CROP;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2379 int i;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2380
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2381 for(i=0; i<h; i++){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2382 dst[0]= cm[(9*(src[0] + src[1]) - (src[-1] + src[2]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2383 dst[1]= cm[(9*(src[1] + src[2]) - (src[ 0] + src[3]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2384 dst[2]= cm[(9*(src[2] + src[3]) - (src[ 1] + src[4]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2385 dst[3]= cm[(9*(src[3] + src[4]) - (src[ 2] + src[5]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2386 dst[4]= cm[(9*(src[4] + src[5]) - (src[ 3] + src[6]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2387 dst[5]= cm[(9*(src[5] + src[6]) - (src[ 4] + src[7]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2388 dst[6]= cm[(9*(src[6] + src[7]) - (src[ 5] + src[8]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2389 dst[7]= cm[(9*(src[7] + src[8]) - (src[ 6] + src[9]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2390 dst+=dstStride;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2391 src+=srcStride;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2392 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2393 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2394
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2395 static void wmv2_mspel8_v_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int w){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2396 uint8_t *cm = cropTbl + MAX_NEG_CROP;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2397 int i;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2398
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2399 for(i=0; i<w; i++){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2400 const int src_1= src[ -srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2401 const int src0 = src[0 ];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2402 const int src1 = src[ srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2403 const int src2 = src[2*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2404 const int src3 = src[3*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2405 const int src4 = src[4*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2406 const int src5 = src[5*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2407 const int src6 = src[6*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2408 const int src7 = src[7*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2409 const int src8 = src[8*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2410 const int src9 = src[9*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2411 dst[0*dstStride]= cm[(9*(src0 + src1) - (src_1 + src2) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2412 dst[1*dstStride]= cm[(9*(src1 + src2) - (src0 + src3) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2413 dst[2*dstStride]= cm[(9*(src2 + src3) - (src1 + src4) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2414 dst[3*dstStride]= cm[(9*(src3 + src4) - (src2 + src5) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2415 dst[4*dstStride]= cm[(9*(src4 + src5) - (src3 + src6) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2416 dst[5*dstStride]= cm[(9*(src5 + src6) - (src4 + src7) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2417 dst[6*dstStride]= cm[(9*(src6 + src7) - (src5 + src8) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2418 dst[7*dstStride]= cm[(9*(src7 + src8) - (src6 + src9) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2419 src++;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2420 dst++;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2421 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2422 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2423
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2424 static void put_mspel8_mc00_c (uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2425 put_pixels8_c(dst, src, stride, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2426 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2427
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2428 static void put_mspel8_mc10_c(uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2429 uint8_t half[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2430 wmv2_mspel8_h_lowpass(half, src, 8, stride, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2431 put_pixels8_l2(dst, src, half, stride, stride, 8, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2432 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2433
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2434 static void put_mspel8_mc20_c(uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2435 wmv2_mspel8_h_lowpass(dst, src, stride, stride, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2436 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2437
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2438 static void put_mspel8_mc30_c(uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2439 uint8_t half[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2440 wmv2_mspel8_h_lowpass(half, src, 8, stride, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2441 put_pixels8_l2(dst, src+1, half, stride, stride, 8, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2442 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2443
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2444 static void put_mspel8_mc02_c(uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2445 wmv2_mspel8_v_lowpass(dst, src, stride, stride, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2446 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2447
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2448 static void put_mspel8_mc12_c(uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2449 uint8_t halfH[88];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2450 uint8_t halfV[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2451 uint8_t halfHV[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2452 wmv2_mspel8_h_lowpass(halfH, src-stride, 8, stride, 11);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2453 wmv2_mspel8_v_lowpass(halfV, src, 8, stride, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2454 wmv2_mspel8_v_lowpass(halfHV, halfH+8, 8, 8, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2455 put_pixels8_l2(dst, halfV, halfHV, stride, 8, 8, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2456 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2457 static void put_mspel8_mc32_c(uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2458 uint8_t halfH[88];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2459 uint8_t halfV[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2460 uint8_t halfHV[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2461 wmv2_mspel8_h_lowpass(halfH, src-stride, 8, stride, 11);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2462 wmv2_mspel8_v_lowpass(halfV, src+1, 8, stride, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2463 wmv2_mspel8_v_lowpass(halfHV, halfH+8, 8, 8, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2464 put_pixels8_l2(dst, halfV, halfHV, stride, 8, 8, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2465 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2466 static void put_mspel8_mc22_c(uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2467 uint8_t halfH[88];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2468 wmv2_mspel8_h_lowpass(halfH, src-stride, 8, stride, 11);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2469 wmv2_mspel8_v_lowpass(dst, halfH+8, stride, 8, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2470 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2471
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2472 static void h263_v_loop_filter_c(uint8_t *src, int stride, int qscale){
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2473 int x;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2474 const int strength= ff_h263_loop_filter_strength[qscale];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2475
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2476 for(x=0; x<8; x++){
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2477 int d1, d2, ad1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2478 int p0= src[x-2*stride];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2479 int p1= src[x-1*stride];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2480 int p2= src[x+0*stride];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2481 int p3= src[x+1*stride];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2482 int d = (p0 - p3 + 4*(p2 - p1)) / 8;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2483
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2484 if (d<-2*strength) d1= 0;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2485 else if(d<- strength) d1=-2*strength - d;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2486 else if(d< strength) d1= d;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2487 else if(d< 2*strength) d1= 2*strength - d;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2488 else d1= 0;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2489
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2490 p1 += d1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2491 p2 -= d1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2492 if(p1&256) p1= ~(p1>>31);
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2493 if(p2&256) p2= ~(p2>>31);
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2494
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2495 src[x-1*stride] = p1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2496 src[x+0*stride] = p2;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2497
1645
michael
parents: 1644
diff changeset
2498 ad1= ABS(d1)>>1;
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2499
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2500 d2= clip((p0-p3)/4, -ad1, ad1);
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2501
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2502 src[x-2*stride] = p0 - d2;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2503 src[x+ stride] = p3 + d2;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2504 }
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2505 }
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2506
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2507 static void h263_h_loop_filter_c(uint8_t *src, int stride, int qscale){
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2508 int y;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2509 const int strength= ff_h263_loop_filter_strength[qscale];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2510
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2511 for(y=0; y<8; y++){
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2512 int d1, d2, ad1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2513 int p0= src[y*stride-2];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2514 int p1= src[y*stride-1];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2515 int p2= src[y*stride+0];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2516 int p3= src[y*stride+1];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2517 int d = (p0 - p3 + 4*(p2 - p1)) / 8;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2518
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2519 if (d<-2*strength) d1= 0;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2520 else if(d<- strength) d1=-2*strength - d;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2521 else if(d< strength) d1= d;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2522 else if(d< 2*strength) d1= 2*strength - d;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2523 else d1= 0;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2524
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2525 p1 += d1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2526 p2 -= d1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2527 if(p1&256) p1= ~(p1>>31);
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2528 if(p2&256) p2= ~(p2>>31);
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2529
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2530 src[y*stride-1] = p1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2531 src[y*stride+0] = p2;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2532
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2533 ad1= ABS(d1)>>1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2534
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2535 d2= clip((p0-p3)/4, -ad1, ad1);
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2536
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2537 src[y*stride-2] = p0 - d2;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2538 src[y*stride+1] = p3 + d2;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2539 }
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2540 }
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2541
2045
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2542 static void h261_loop_filter_c(uint8_t *src, int stride){
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2543 int x,y,xy,yz;
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2544 int temp[64];
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2545
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2546 for(x=0; x<8; x++){
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2547 temp[x ] = 4*src[x ];
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2548 temp[x + 7*8] = 4*src[x + 7*stride];
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2549 }
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2550 for(y=1; y<7; y++){
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2551 for(x=0; x<8; x++){
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2552 xy = y * stride + x;
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2553 yz = y * 8 + x;
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2554 temp[yz] = src[xy - stride] + 2*src[xy] + src[xy + stride];
2044
b6f2add2511e h261 decoder by (Maarten Daniels <maarten.daniels at student dot luc dot ac dot be>)
michael
parents: 1984
diff changeset
2555 }
b6f2add2511e h261 decoder by (Maarten Daniels <maarten.daniels at student dot luc dot ac dot be>)
michael
parents: 1984
diff changeset
2556 }
2045
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2557
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2558 for(y=0; y<8; y++){
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2559 src[ y*stride] = (temp[ y*8] + 2)>>2;
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2560 src[7+y*stride] = (temp[7+y*8] + 2)>>2;
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2561 for(x=1; x<7; x++){
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2562 xy = y * stride + x;
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2563 yz = y * 8 + x;
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2564 src[xy] = (temp[yz-1] + 2*temp[yz] + temp[yz+1] + 8)>>4;
2044
b6f2add2511e h261 decoder by (Maarten Daniels <maarten.daniels at student dot luc dot ac dot be>)
michael
parents: 1984
diff changeset
2565 }
b6f2add2511e h261 decoder by (Maarten Daniels <maarten.daniels at student dot luc dot ac dot be>)
michael
parents: 1984
diff changeset
2566 }
b6f2add2511e h261 decoder by (Maarten Daniels <maarten.daniels at student dot luc dot ac dot be>)
michael
parents: 1984
diff changeset
2567 }
b6f2add2511e h261 decoder by (Maarten Daniels <maarten.daniels at student dot luc dot ac dot be>)
michael
parents: 1984
diff changeset
2568
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2569 static inline int pix_abs16_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
2570 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
2571 int s, i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2572
986e461dc072 Initial revision
glantau
parents:
diff changeset
2573 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2574 for(i=0;i<h;i++) {
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
2575 s += abs(pix1[0] - pix2[0]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2576 s += abs(pix1[1] - pix2[1]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2577 s += abs(pix1[2] - pix2[2]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2578 s += abs(pix1[3] - pix2[3]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2579 s += abs(pix1[4] - pix2[4]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2580 s += abs(pix1[5] - pix2[5]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2581 s += abs(pix1[6] - pix2[6]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2582 s += abs(pix1[7] - pix2[7]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2583 s += abs(pix1[8] - pix2[8]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2584 s += abs(pix1[9] - pix2[9]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2585 s += abs(pix1[10] - pix2[10]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2586 s += abs(pix1[11] - pix2[11]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2587 s += abs(pix1[12] - pix2[12]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2588 s += abs(pix1[13] - pix2[13]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2589 s += abs(pix1[14] - pix2[14]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2590 s += abs(pix1[15] - pix2[15]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2591 pix1 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2592 pix2 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2593 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
2594 return s;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2595 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
2596
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2597 static int pix_abs16_x2_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
2598 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
2599 int s, i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2600
986e461dc072 Initial revision
glantau
parents:
diff changeset
2601 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2602 for(i=0;i<h;i++) {
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
2603 s += abs(pix1[0] - avg2(pix2[0], pix2[1]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2604 s += abs(pix1[1] - avg2(pix2[1], pix2[2]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2605 s += abs(pix1[2] - avg2(pix2[2], pix2[3]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2606 s += abs(pix1[3] - avg2(pix2[3], pix2[4]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2607 s += abs(pix1[4] - avg2(pix2[4], pix2[5]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2608 s += abs(pix1[5] - avg2(pix2[5], pix2[6]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2609 s += abs(pix1[6] - avg2(pix2[6], pix2[7]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2610 s += abs(pix1[7] - avg2(pix2[7], pix2[8]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2611 s += abs(pix1[8] - avg2(pix2[8], pix2[9]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2612 s += abs(pix1[9] - avg2(pix2[9], pix2[10]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2613 s += abs(pix1[10] - avg2(pix2[10], pix2[11]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2614 s += abs(pix1[11] - avg2(pix2[11], pix2[12]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2615 s += abs(pix1[12] - avg2(pix2[12], pix2[13]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2616 s += abs(pix1[13] - avg2(pix2[13], pix2[14]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2617 s += abs(pix1[14] - avg2(pix2[14], pix2[15]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2618 s += abs(pix1[15] - avg2(pix2[15], pix2[16]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2619 pix1 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2620 pix2 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2621 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
2622 return s;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2623 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
2624
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2625 static int pix_abs16_y2_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
2626 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
2627 int s, i;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2628 uint8_t *pix3 = pix2 + line_size;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
2629
986e461dc072 Initial revision
glantau
parents:
diff changeset
2630 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2631 for(i=0;i<h;i++) {
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
2632 s += abs(pix1[0] - avg2(pix2[0], pix3[0]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2633 s += abs(pix1[1] - avg2(pix2[1], pix3[1]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2634 s += abs(pix1[2] - avg2(pix2[2], pix3[2]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2635 s += abs(pix1[3] - avg2(pix2[3], pix3[3]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2636 s += abs(pix1[4] - avg2(pix2[4], pix3[4]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2637 s += abs(pix1[5] - avg2(pix2[5], pix3[5]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2638 s += abs(pix1[6] - avg2(pix2[6], pix3[6]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2639 s += abs(pix1[7] - avg2(pix2[7], pix3[7]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2640 s += abs(pix1[8] - avg2(pix2[8], pix3[8]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2641 s += abs(pix1[9] - avg2(pix2[9], pix3[9]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2642 s += abs(pix1[10] - avg2(pix2[10], pix3[10]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2643 s += abs(pix1[11] - avg2(pix2[11], pix3[11]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2644 s += abs(pix1[12] - avg2(pix2[12], pix3[12]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2645 s += abs(pix1[13] - avg2(pix2[13], pix3[13]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2646 s += abs(pix1[14] - avg2(pix2[14], pix3[14]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2647 s += abs(pix1[15] - avg2(pix2[15], pix3[15]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2648 pix1 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2649 pix2 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2650 pix3 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2651 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
2652 return s;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2653 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
2654
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2655 static int pix_abs16_xy2_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
2656 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
2657 int s, i;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2658 uint8_t *pix3 = pix2 + line_size;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
2659
986e461dc072 Initial revision
glantau
parents:
diff changeset
2660 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2661 for(i=0;i<h;i++) {
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
2662 s += abs(pix1[0] - avg4(pix2[0], pix2[1], pix3[0], pix3[1]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2663 s += abs(pix1[1] - avg4(pix2[1], pix2[2], pix3[1], pix3[2]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2664 s += abs(pix1[2] - avg4(pix2[2], pix2[3], pix3[2], pix3[3]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2665 s += abs(pix1[3] - avg4(pix2[3], pix2[4], pix3[3], pix3[4]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2666 s += abs(pix1[4] - avg4(pix2[4], pix2[5], pix3[4], pix3[5]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2667 s += abs(pix1[5] - avg4(pix2[5], pix2[6], pix3[5], pix3[6]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2668 s += abs(pix1[6] - avg4(pix2[6], pix2[7], pix3[6], pix3[7]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2669 s += abs(pix1[7] - avg4(pix2[7], pix2[8], pix3[7], pix3[8]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2670 s += abs(pix1[8] - avg4(pix2[8], pix2[9], pix3[8], pix3[9]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2671 s += abs(pix1[9] - avg4(pix2[9], pix2[10], pix3[9], pix3[10]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2672 s += abs(pix1[10] - avg4(pix2[10], pix2[11], pix3[10], pix3[11]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2673 s += abs(pix1[11] - avg4(pix2[11], pix2[12], pix3[11], pix3[12]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2674 s += abs(pix1[12] - avg4(pix2[12], pix2[13], pix3[12], pix3[13]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2675 s += abs(pix1[13] - avg4(pix2[13], pix2[14], pix3[13], pix3[14]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2676 s += abs(pix1[14] - avg4(pix2[14], pix2[15], pix3[14], pix3[15]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2677 s += abs(pix1[15] - avg4(pix2[15], pix2[16], pix3[15], pix3[16]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2678 pix1 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2679 pix2 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2680 pix3 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2681 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
2682 return s;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2683 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
2684
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2685 static inline int pix_abs8_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2686 {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2687 int s, i;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2688
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2689 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2690 for(i=0;i<h;i++) {
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2691 s += abs(pix1[0] - pix2[0]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2692 s += abs(pix1[1] - pix2[1]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2693 s += abs(pix1[2] - pix2[2]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2694 s += abs(pix1[3] - pix2[3]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2695 s += abs(pix1[4] - pix2[4]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2696 s += abs(pix1[5] - pix2[5]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2697 s += abs(pix1[6] - pix2[6]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2698 s += abs(pix1[7] - pix2[7]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2699 pix1 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2700 pix2 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2701 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2702 return s;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2703 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2704
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2705 static int pix_abs8_x2_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2706 {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2707 int s, i;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2708
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2709 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2710 for(i=0;i<h;i++) {
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2711 s += abs(pix1[0] - avg2(pix2[0], pix2[1]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2712 s += abs(pix1[1] - avg2(pix2[1], pix2[2]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2713 s += abs(pix1[2] - avg2(pix2[2], pix2[3]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2714 s += abs(pix1[3] - avg2(pix2[3], pix2[4]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2715 s += abs(pix1[4] - avg2(pix2[4], pix2[5]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2716 s += abs(pix1[5] - avg2(pix2[5], pix2[6]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2717 s += abs(pix1[6] - avg2(pix2[6], pix2[7]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2718 s += abs(pix1[7] - avg2(pix2[7], pix2[8]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2719 pix1 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2720 pix2 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2721 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2722 return s;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2723 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2724
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2725 static int pix_abs8_y2_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2726 {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2727 int s, i;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2728 uint8_t *pix3 = pix2 + line_size;
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2729
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2730 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2731 for(i=0;i<h;i++) {
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2732 s += abs(pix1[0] - avg2(pix2[0], pix3[0]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2733 s += abs(pix1[1] - avg2(pix2[1], pix3[1]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2734 s += abs(pix1[2] - avg2(pix2[2], pix3[2]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2735 s += abs(pix1[3] - avg2(pix2[3], pix3[3]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2736 s += abs(pix1[4] - avg2(pix2[4], pix3[4]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2737 s += abs(pix1[5] - avg2(pix2[5], pix3[5]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2738 s += abs(pix1[6] - avg2(pix2[6], pix3[6]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2739 s += abs(pix1[7] - avg2(pix2[7], pix3[7]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2740 pix1 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2741 pix2 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2742 pix3 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2743 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2744 return s;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2745 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2746
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2747 static int pix_abs8_xy2_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2748 {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2749 int s, i;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2750 uint8_t *pix3 = pix2 + line_size;
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2751
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2752 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2753 for(i=0;i<h;i++) {
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2754 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
2755 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
2756 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
2757 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
2758 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
2759 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
2760 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
2761 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
2762 pix1 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2763 pix2 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2764 pix3 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2765 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2766 return s;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2767 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
2768
2066
4bfb146e701b nsse weight
michael
parents: 2065
diff changeset
2769 static int nsse16_c(MpegEncContext *c, uint8_t *s1, uint8_t *s2, int stride, int h){
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2770 int score1=0;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2771 int score2=0;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2772 int x,y;
2066
4bfb146e701b nsse weight
michael
parents: 2065
diff changeset
2773
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2774 for(y=0; y<h; y++){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2775 for(x=0; x<16; x++){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2776 score1+= (s1[x ] - s2[x ])*(s1[x ] - s2[x ]);
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2777 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2778 if(y+1<h){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2779 for(x=0; x<15; x++){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2780 score2+= ABS( s1[x ] - s1[x +stride]
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2781 - s1[x+1] + s1[x+1+stride])
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2782 -ABS( s2[x ] - s2[x +stride]
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2783 - s2[x+1] + s2[x+1+stride]);
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2784 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2785 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2786 s1+= stride;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2787 s2+= stride;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2788 }
2066
4bfb146e701b nsse weight
michael
parents: 2065
diff changeset
2789
4bfb146e701b nsse weight
michael
parents: 2065
diff changeset
2790 if(c) return score1 + ABS(score2)*c->avctx->nsse_weight;
4bfb146e701b nsse weight
michael
parents: 2065
diff changeset
2791 else return score1 + ABS(score2)*8;
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2792 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2793
2066
4bfb146e701b nsse weight
michael
parents: 2065
diff changeset
2794 static int nsse8_c(MpegEncContext *c, uint8_t *s1, uint8_t *s2, int stride, int h){
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2795 int score1=0;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2796 int score2=0;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2797 int x,y;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2798
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2799 for(y=0; y<h; y++){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2800 for(x=0; x<8; x++){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2801 score1+= (s1[x ] - s2[x ])*(s1[x ] - s2[x ]);
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2802 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2803 if(y+1<h){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2804 for(x=0; x<7; x++){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2805 score2+= ABS( s1[x ] - s1[x +stride]
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2806 - s1[x+1] + s1[x+1+stride])
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2807 -ABS( s2[x ] - s2[x +stride]
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2808 - s2[x+1] + s2[x+1+stride]);
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2809 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2810 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2811 s1+= stride;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2812 s2+= stride;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2813 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2814
2066
4bfb146e701b nsse weight
michael
parents: 2065
diff changeset
2815 if(c) return score1 + ABS(score2)*c->avctx->nsse_weight;
4bfb146e701b nsse weight
michael
parents: 2065
diff changeset
2816 else return score1 + ABS(score2)*8;
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2817 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2818
1784
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2819 static int try_8x8basis_c(int16_t rem[64], int16_t weight[64], int16_t basis[64], int scale){
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2820 int i;
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2821 unsigned int sum=0;
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2822
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2823 for(i=0; i<8*8; i++){
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2824 int b= rem[i] + ((basis[i]*scale + (1<<(BASIS_SHIFT - RECON_SHIFT-1)))>>(BASIS_SHIFT - RECON_SHIFT));
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2825 int w= weight[i];
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2826 b>>= RECON_SHIFT;
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2827 assert(-512<b && b<512);
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2828
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2829 sum += (w*b)*(w*b)>>4;
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2830 }
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2831 return sum>>2;
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2832 }
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2833
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2834 static void add_8x8basis_c(int16_t rem[64], int16_t basis[64], int scale){
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2835 int i;
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2836
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2837 for(i=0; i<8*8; i++){
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2838 rem[i] += (basis[i]*scale + (1<<(BASIS_SHIFT - RECON_SHIFT-1)))>>(BASIS_SHIFT - RECON_SHIFT);
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2839 }
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2840 }
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
2841
1100
207f27932628 ff_block_permute() dox
michaelni
parents: 1092
diff changeset
2842 /**
207f27932628 ff_block_permute() dox
michaelni
parents: 1092
diff changeset
2843 * permutes an 8x8 block.
1101
a3a8017dee70 more idiot proof dox
michaelni
parents: 1100
diff changeset
2844 * @param block the block which will be permuted according to the given permutation vector
1100
207f27932628 ff_block_permute() dox
michaelni
parents: 1092
diff changeset
2845 * @param permutation the permutation vector
207f27932628 ff_block_permute() dox
michaelni
parents: 1092
diff changeset
2846 * @param last the last non zero coefficient in scantable order, used to speed the permutation up
1101
a3a8017dee70 more idiot proof dox
michaelni
parents: 1100
diff changeset
2847 * @param scantable the used scantable, this is only used to speed the permutation up, the block is not
a3a8017dee70 more idiot proof dox
michaelni
parents: 1100
diff changeset
2848 * (inverse) permutated to scantable order!
1100
207f27932628 ff_block_permute() dox
michaelni
parents: 1092
diff changeset
2849 */
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2850 void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *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
2851 {
764
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
2852 int i;
945
463f7260b155 trellis quantization
michaelni
parents: 936
diff changeset
2853 DCTELEM temp[64];
764
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
2854
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
2855 if(last<=0) return;
882
34943cb0c2fb * remaining part of John Ryland's patch
kabi
parents: 866
diff changeset
2856 //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
2857
764
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
2858 for(i=0; i<=last; i++){
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
2859 const int j= scantable[i];
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
2860 temp[j]= block[j];
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
2861 block[j]=0;
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
2862 }
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
2863
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
2864 for(i=0; i<=last; i++){
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
2865 const int j= scantable[i];
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
2866 const int perm_j= permutation[j];
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
2867 block[perm_j]= temp[j];
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
2868 }
174
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
2869 }
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
2870
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2871 static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2872 return 0;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2873 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2874
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2875 void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2876 int i;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2877
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2878 memset(cmp, 0, sizeof(void*)*5);
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2879
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2880 for(i=0; i<5; i++){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2881 switch(type&0xFF){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2882 case FF_CMP_SAD:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2883 cmp[i]= c->sad[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2884 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2885 case FF_CMP_SATD:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2886 cmp[i]= c->hadamard8_diff[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2887 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2888 case FF_CMP_SSE:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2889 cmp[i]= c->sse[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2890 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2891 case FF_CMP_DCT:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2892 cmp[i]= c->dct_sad[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2893 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2894 case FF_CMP_PSNR:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2895 cmp[i]= c->quant_psnr[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2896 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2897 case FF_CMP_BIT:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2898 cmp[i]= c->bit[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2899 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2900 case FF_CMP_RD:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2901 cmp[i]= c->rd[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2902 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2903 case FF_CMP_VSAD:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2904 cmp[i]= c->vsad[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2905 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2906 case FF_CMP_VSSE:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2907 cmp[i]= c->vsse[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2908 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2909 case FF_CMP_ZERO:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2910 cmp[i]= zero_cmp;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2911 break;
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2912 case FF_CMP_NSSE:
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2913 cmp[i]= c->nsse[i];
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
2914 break;
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
2915 case FF_CMP_W53:
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
2916 cmp[i]= c->w53[i];
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
2917 break;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
2918 case FF_CMP_W97:
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
2919 cmp[i]= c->w97[i];
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
2920 break;
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2921 default:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2922 av_log(NULL, AV_LOG_ERROR,"internal error in cmp function selection\n");
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2923 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2924 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2925 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
2926
1101
a3a8017dee70 more idiot proof dox
michaelni
parents: 1100
diff changeset
2927 /**
a3a8017dee70 more idiot proof dox
michaelni
parents: 1100
diff changeset
2928 * memset(blocks, 0, sizeof(DCTELEM)*6*64)
a3a8017dee70 more idiot proof dox
michaelni
parents: 1100
diff changeset
2929 */
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
2930 static void clear_blocks_c(DCTELEM *blocks)
296
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
2931 {
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
2932 memset(blocks, 0, sizeof(DCTELEM)*6*64);
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
2933 }
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
2934
866
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2935 static void add_bytes_c(uint8_t *dst, uint8_t *src, int w){
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2936 int i;
996
ad44196ea483 add/diff_bytes bugfix patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>)
michaelni
parents: 984
diff changeset
2937 for(i=0; i+7<w; i+=8){
866
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2938 dst[i+0] += src[i+0];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2939 dst[i+1] += src[i+1];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2940 dst[i+2] += src[i+2];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2941 dst[i+3] += src[i+3];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2942 dst[i+4] += src[i+4];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2943 dst[i+5] += src[i+5];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2944 dst[i+6] += src[i+6];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2945 dst[i+7] += src[i+7];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2946 }
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2947 for(; i<w; i++)
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2948 dst[i+0] += src[i+0];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2949 }
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2950
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2951 static void diff_bytes_c(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w){
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2952 int i;
996
ad44196ea483 add/diff_bytes bugfix patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>)
michaelni
parents: 984
diff changeset
2953 for(i=0; i+7<w; i+=8){
866
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2954 dst[i+0] = src1[i+0]-src2[i+0];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2955 dst[i+1] = src1[i+1]-src2[i+1];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2956 dst[i+2] = src1[i+2]-src2[i+2];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2957 dst[i+3] = src1[i+3]-src2[i+3];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2958 dst[i+4] = src1[i+4]-src2[i+4];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2959 dst[i+5] = src1[i+5]-src2[i+5];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2960 dst[i+6] = src1[i+6]-src2[i+6];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2961 dst[i+7] = src1[i+7]-src2[i+7];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2962 }
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2963 for(; i<w; i++)
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2964 dst[i+0] = src1[i+0]-src2[i+0];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2965 }
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
2966
1527
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
2967 static void sub_hfyu_median_prediction_c(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top){
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
2968 int i;
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
2969 uint8_t l, lt;
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
2970
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
2971 l= *left;
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
2972 lt= *left_top;
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
2973
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
2974 for(i=0; i<w; i++){
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
2975 const int pred= mid_pred(l, src1[i], (l + src1[i] - lt)&0xFF);
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
2976 lt= src1[i];
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
2977 l= src2[i];
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
2978 dst[i]= l - pred;
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
2979 }
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
2980
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
2981 *left= l;
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
2982 *left_top= lt;
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
2983 }
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
2984
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2985 #define BUTTERFLY2(o1,o2,i1,i2) \
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2986 o1= (i1)+(i2);\
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2987 o2= (i1)-(i2);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2988
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2989 #define BUTTERFLY1(x,y) \
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2990 {\
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2991 int a,b;\
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2992 a= x;\
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2993 b= y;\
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2994 x= a+b;\
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2995 y= a-b;\
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2996 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2997
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2998 #define BUTTERFLYA(x,y) (ABS((x)+(y)) + ABS((x)-(y)))
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2999
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3000 static int hadamard8_diff8x8_c(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h){
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3001 int i;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3002 int temp[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3003 int sum=0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3004
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3005 assert(h==8);
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3006
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3007 for(i=0; i<8; i++){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3008 //FIXME try pointer walks
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3009 BUTTERFLY2(temp[8*i+0], temp[8*i+1], src[stride*i+0]-dst[stride*i+0],src[stride*i+1]-dst[stride*i+1]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3010 BUTTERFLY2(temp[8*i+2], temp[8*i+3], src[stride*i+2]-dst[stride*i+2],src[stride*i+3]-dst[stride*i+3]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3011 BUTTERFLY2(temp[8*i+4], temp[8*i+5], src[stride*i+4]-dst[stride*i+4],src[stride*i+5]-dst[stride*i+5]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3012 BUTTERFLY2(temp[8*i+6], temp[8*i+7], src[stride*i+6]-dst[stride*i+6],src[stride*i+7]-dst[stride*i+7]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3013
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3014 BUTTERFLY1(temp[8*i+0], temp[8*i+2]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3015 BUTTERFLY1(temp[8*i+1], temp[8*i+3]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3016 BUTTERFLY1(temp[8*i+4], temp[8*i+6]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3017 BUTTERFLY1(temp[8*i+5], temp[8*i+7]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3018
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3019 BUTTERFLY1(temp[8*i+0], temp[8*i+4]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3020 BUTTERFLY1(temp[8*i+1], temp[8*i+5]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3021 BUTTERFLY1(temp[8*i+2], temp[8*i+6]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3022 BUTTERFLY1(temp[8*i+3], temp[8*i+7]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3023 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3024
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3025 for(i=0; i<8; i++){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3026 BUTTERFLY1(temp[8*0+i], temp[8*1+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3027 BUTTERFLY1(temp[8*2+i], temp[8*3+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3028 BUTTERFLY1(temp[8*4+i], temp[8*5+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3029 BUTTERFLY1(temp[8*6+i], temp[8*7+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3030
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3031 BUTTERFLY1(temp[8*0+i], temp[8*2+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3032 BUTTERFLY1(temp[8*1+i], temp[8*3+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3033 BUTTERFLY1(temp[8*4+i], temp[8*6+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3034 BUTTERFLY1(temp[8*5+i], temp[8*7+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3035
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3036 sum +=
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3037 BUTTERFLYA(temp[8*0+i], temp[8*4+i])
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3038 +BUTTERFLYA(temp[8*1+i], temp[8*5+i])
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3039 +BUTTERFLYA(temp[8*2+i], temp[8*6+i])
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3040 +BUTTERFLYA(temp[8*3+i], temp[8*7+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3041 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3042 #if 0
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3043 static int maxi=0;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3044 if(sum>maxi){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3045 maxi=sum;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3046 printf("MAX:%d\n", maxi);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3047 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3048 #endif
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3049 return sum;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3050 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3051
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3052 static int hadamard8_intra8x8_c(/*MpegEncContext*/ void *s, uint8_t *src, uint8_t *dummy, int stride, int h){
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3053 int i;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3054 int temp[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3055 int sum=0;
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3056
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3057 assert(h==8);
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3058
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3059 for(i=0; i<8; i++){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3060 //FIXME try pointer walks
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3061 BUTTERFLY2(temp[8*i+0], temp[8*i+1], src[stride*i+0],src[stride*i+1]);
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3062 BUTTERFLY2(temp[8*i+2], temp[8*i+3], src[stride*i+2],src[stride*i+3]);
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3063 BUTTERFLY2(temp[8*i+4], temp[8*i+5], src[stride*i+4],src[stride*i+5]);
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3064 BUTTERFLY2(temp[8*i+6], temp[8*i+7], src[stride*i+6],src[stride*i+7]);
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3065
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3066 BUTTERFLY1(temp[8*i+0], temp[8*i+2]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3067 BUTTERFLY1(temp[8*i+1], temp[8*i+3]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3068 BUTTERFLY1(temp[8*i+4], temp[8*i+6]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3069 BUTTERFLY1(temp[8*i+5], temp[8*i+7]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3070
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3071 BUTTERFLY1(temp[8*i+0], temp[8*i+4]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3072 BUTTERFLY1(temp[8*i+1], temp[8*i+5]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3073 BUTTERFLY1(temp[8*i+2], temp[8*i+6]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3074 BUTTERFLY1(temp[8*i+3], temp[8*i+7]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3075 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3076
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3077 for(i=0; i<8; i++){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3078 BUTTERFLY1(temp[8*0+i], temp[8*1+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3079 BUTTERFLY1(temp[8*2+i], temp[8*3+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3080 BUTTERFLY1(temp[8*4+i], temp[8*5+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3081 BUTTERFLY1(temp[8*6+i], temp[8*7+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3082
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3083 BUTTERFLY1(temp[8*0+i], temp[8*2+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3084 BUTTERFLY1(temp[8*1+i], temp[8*3+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3085 BUTTERFLY1(temp[8*4+i], temp[8*6+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3086 BUTTERFLY1(temp[8*5+i], temp[8*7+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3087
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3088 sum +=
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3089 BUTTERFLYA(temp[8*0+i], temp[8*4+i])
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3090 +BUTTERFLYA(temp[8*1+i], temp[8*5+i])
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3091 +BUTTERFLYA(temp[8*2+i], temp[8*6+i])
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3092 +BUTTERFLYA(temp[8*3+i], temp[8*7+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3093 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3094
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3095 sum -= ABS(temp[8*0] + temp[8*4]); // -mean
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3096
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3097 return sum;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3098 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3099
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3100 static int dct_sad8x8_c(/*MpegEncContext*/ void *c, uint8_t *src1, uint8_t *src2, int stride, int h){
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3101 MpegEncContext * const s= (MpegEncContext *)c;
1016
michaelni
parents: 1013
diff changeset
3102 uint64_t __align8 aligned_temp[sizeof(DCTELEM)*64/8];
michaelni
parents: 1013
diff changeset
3103 DCTELEM * const temp= (DCTELEM*)aligned_temp;
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3104 int sum=0, i;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3105
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3106 assert(h==8);
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3107
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3108 s->dsp.diff_pixels(temp, src1, src2, stride);
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3109 s->dsp.fdct(temp);
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3110
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3111 for(i=0; i<64; i++)
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3112 sum+= ABS(temp[i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3113
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3114 return sum;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3115 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3116
1008
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 1007
diff changeset
3117 void simple_idct(DCTELEM *block); //FIXME
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3118
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3119 static int quant_psnr8x8_c(/*MpegEncContext*/ void *c, uint8_t *src1, uint8_t *src2, int stride, int h){
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3120 MpegEncContext * const s= (MpegEncContext *)c;
1016
michaelni
parents: 1013
diff changeset
3121 uint64_t __align8 aligned_temp[sizeof(DCTELEM)*64*2/8];
michaelni
parents: 1013
diff changeset
3122 DCTELEM * const temp= (DCTELEM*)aligned_temp;
michaelni
parents: 1013
diff changeset
3123 DCTELEM * const bak = ((DCTELEM*)aligned_temp)+64;
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3124 int sum=0, i;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3125
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3126 assert(h==8);
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3127 s->mb_intra=0;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3128
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3129 s->dsp.diff_pixels(temp, src1, src2, stride);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3130
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3131 memcpy(bak, temp, 64*sizeof(DCTELEM));
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3132
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3133 s->block_last_index[0/*FIXME*/]= s->fast_dct_quantize(s, temp, 0/*FIXME*/, s->qscale, &i);
1689
1a2db2073848 split intra / inter dequantization
michael
parents: 1645
diff changeset
3134 s->dct_unquantize_inter(s, temp, 0, s->qscale);
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3135 simple_idct(temp); //FIXME
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3136
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3137 for(i=0; i<64; i++)
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3138 sum+= (temp[i]-bak[i])*(temp[i]-bak[i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3139
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3140 return sum;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3141 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3142
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3143 static int rd8x8_c(/*MpegEncContext*/ void *c, uint8_t *src1, uint8_t *src2, int stride, int h){
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3144 MpegEncContext * const s= (MpegEncContext *)c;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
3145 const uint8_t *scantable= s->intra_scantable.permutated;
1016
michaelni
parents: 1013
diff changeset
3146 uint64_t __align8 aligned_temp[sizeof(DCTELEM)*64/8];
michaelni
parents: 1013
diff changeset
3147 uint64_t __align8 aligned_bak[stride];
michaelni
parents: 1013
diff changeset
3148 DCTELEM * const temp= (DCTELEM*)aligned_temp;
michaelni
parents: 1013
diff changeset
3149 uint8_t * const bak= (uint8_t*)aligned_bak;
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3150 int i, last, run, bits, level, distoration, start_i;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3151 const int esc_length= s->ac_esc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3152 uint8_t * length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3153 uint8_t * last_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3154
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3155 assert(h==8);
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3156
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3157 for(i=0; i<8; i++){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3158 ((uint32_t*)(bak + i*stride))[0]= ((uint32_t*)(src2 + i*stride))[0];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3159 ((uint32_t*)(bak + i*stride))[1]= ((uint32_t*)(src2 + i*stride))[1];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3160 }
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3161
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3162 s->dsp.diff_pixels(temp, src1, src2, stride);
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3163
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3164 s->block_last_index[0/*FIXME*/]= last= s->fast_dct_quantize(s, temp, 0/*FIXME*/, s->qscale, &i);
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3165
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3166 bits=0;
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3167
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3168 if (s->mb_intra) {
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3169 start_i = 1;
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3170 length = s->intra_ac_vlc_length;
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3171 last_length= s->intra_ac_vlc_last_length;
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3172 bits+= s->luma_dc_vlc_length[temp[0] + 256]; //FIXME chroma
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3173 } else {
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3174 start_i = 0;
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3175 length = s->inter_ac_vlc_length;
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3176 last_length= s->inter_ac_vlc_last_length;
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3177 }
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3178
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3179 if(last>=start_i){
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3180 run=0;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3181 for(i=start_i; i<last; i++){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3182 int j= scantable[i];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3183 level= temp[j];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3184
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3185 if(level){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3186 level+=64;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3187 if((level&(~127)) == 0){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3188 bits+= length[UNI_AC_ENC_INDEX(run, level)];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3189 }else
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3190 bits+= esc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3191 run=0;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3192 }else
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3193 run++;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3194 }
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3195 i= scantable[last];
1011
michaelni
parents: 1008
diff changeset
3196
michaelni
parents: 1008
diff changeset
3197 level= temp[i] + 64;
michaelni
parents: 1008
diff changeset
3198
michaelni
parents: 1008
diff changeset
3199 assert(level - 64);
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3200
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3201 if((level&(~127)) == 0){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3202 bits+= last_length[UNI_AC_ENC_INDEX(run, level)];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3203 }else
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3204 bits+= esc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3205
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3206 }
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3207
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3208 if(last>=0){
1689
1a2db2073848 split intra / inter dequantization
michael
parents: 1645
diff changeset
3209 if(s->mb_intra)
1a2db2073848 split intra / inter dequantization
michael
parents: 1645
diff changeset
3210 s->dct_unquantize_intra(s, temp, 0, s->qscale);
1a2db2073848 split intra / inter dequantization
michael
parents: 1645
diff changeset
3211 else
1a2db2073848 split intra / inter dequantization
michael
parents: 1645
diff changeset
3212 s->dct_unquantize_inter(s, temp, 0, s->qscale);
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3213 }
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3214
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3215 s->dsp.idct_add(bak, stride, temp);
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3216
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3217 distoration= s->dsp.sse[1](NULL, bak, src1, stride, 8);
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3218
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3219 return distoration + ((bits*s->qscale*s->qscale*109 + 64)>>7);
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3220 }
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3221
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3222 static int bit8x8_c(/*MpegEncContext*/ void *c, uint8_t *src1, uint8_t *src2, int stride, int h){
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3223 MpegEncContext * const s= (MpegEncContext *)c;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
3224 const uint8_t *scantable= s->intra_scantable.permutated;
1016
michaelni
parents: 1013
diff changeset
3225 uint64_t __align8 aligned_temp[sizeof(DCTELEM)*64/8];
michaelni
parents: 1013
diff changeset
3226 DCTELEM * const temp= (DCTELEM*)aligned_temp;
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3227 int i, last, run, bits, level, start_i;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3228 const int esc_length= s->ac_esc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3229 uint8_t * length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3230 uint8_t * last_length;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3231
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3232 assert(h==8);
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3233
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3234 s->dsp.diff_pixels(temp, src1, src2, stride);
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3235
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3236 s->block_last_index[0/*FIXME*/]= last= s->fast_dct_quantize(s, temp, 0/*FIXME*/, s->qscale, &i);
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3237
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3238 bits=0;
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3239
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3240 if (s->mb_intra) {
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3241 start_i = 1;
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3242 length = s->intra_ac_vlc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3243 last_length= s->intra_ac_vlc_last_length;
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3244 bits+= s->luma_dc_vlc_length[temp[0] + 256]; //FIXME chroma
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3245 } else {
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3246 start_i = 0;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3247 length = s->inter_ac_vlc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3248 last_length= s->inter_ac_vlc_last_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3249 }
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3250
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3251 if(last>=start_i){
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3252 run=0;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3253 for(i=start_i; i<last; i++){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3254 int j= scantable[i];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3255 level= temp[j];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3256
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3257 if(level){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3258 level+=64;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3259 if((level&(~127)) == 0){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3260 bits+= length[UNI_AC_ENC_INDEX(run, level)];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3261 }else
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3262 bits+= esc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3263 run=0;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3264 }else
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3265 run++;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3266 }
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3267 i= scantable[last];
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3268
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3269 level= temp[i] + 64;
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3270
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3271 assert(level - 64);
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3272
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3273 if((level&(~127)) == 0){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3274 bits+= last_length[UNI_AC_ENC_INDEX(run, level)];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3275 }else
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3276 bits+= esc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3277 }
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3278
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3279 return bits;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3280 }
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3281
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3282 static int vsad_intra16_c(/*MpegEncContext*/ void *c, uint8_t *s, uint8_t *dummy, int stride, int h){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3283 int score=0;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3284 int x,y;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3285
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3286 for(y=1; y<h; y++){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3287 for(x=0; x<16; x+=4){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3288 score+= ABS(s[x ] - s[x +stride]) + ABS(s[x+1] - s[x+1+stride])
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3289 +ABS(s[x+2] - s[x+2+stride]) + ABS(s[x+3] - s[x+3+stride]);
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3290 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3291 s+= stride;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3292 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3293
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3294 return score;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3295 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3296
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3297 static int vsad16_c(/*MpegEncContext*/ void *c, uint8_t *s1, uint8_t *s2, int stride, int h){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3298 int score=0;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3299 int x,y;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3300
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3301 for(y=1; y<h; y++){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3302 for(x=0; x<16; x++){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3303 score+= ABS(s1[x ] - s2[x ] - s1[x +stride] + s2[x +stride]);
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3304 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3305 s1+= stride;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3306 s2+= stride;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3307 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3308
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3309 return score;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3310 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3311
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3312 #define SQ(a) ((a)*(a))
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3313 static int vsse_intra16_c(/*MpegEncContext*/ void *c, uint8_t *s, uint8_t *dummy, int stride, int h){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3314 int score=0;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3315 int x,y;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3316
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3317 for(y=1; y<h; y++){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3318 for(x=0; x<16; x+=4){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3319 score+= SQ(s[x ] - s[x +stride]) + SQ(s[x+1] - s[x+1+stride])
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3320 +SQ(s[x+2] - s[x+2+stride]) + SQ(s[x+3] - s[x+3+stride]);
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3321 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3322 s+= stride;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3323 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3324
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3325 return score;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3326 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3327
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3328 static int vsse16_c(/*MpegEncContext*/ void *c, uint8_t *s1, uint8_t *s2, int stride, int h){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3329 int score=0;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3330 int x,y;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3331
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3332 for(y=1; y<h; y++){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3333 for(x=0; x<16; x++){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3334 score+= SQ(s1[x ] - s2[x ] - s1[x +stride] + s2[x +stride]);
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3335 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3336 s1+= stride;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3337 s2+= stride;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3338 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3339
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3340 return score;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3341 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3342
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3343 WARPER8_16_SQ(hadamard8_diff8x8_c, hadamard8_diff16_c)
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3344 WARPER8_16_SQ(hadamard8_intra8x8_c, hadamard8_intra16_c)
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3345 WARPER8_16_SQ(dct_sad8x8_c, dct_sad16_c)
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3346 WARPER8_16_SQ(quant_psnr8x8_c, quant_psnr16_c)
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3347 WARPER8_16_SQ(rd8x8_c, rd16_c)
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3348 WARPER8_16_SQ(bit8x8_c, bit16_c)
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3349
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3350 /* XXX: those functions should be suppressed ASAP when all IDCTs are
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3351 converted */
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3352 static void ff_jref_idct_put(uint8_t *dest, int line_size, DCTELEM *block)
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3353 {
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3354 j_rev_dct (block);
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3355 put_pixels_clamped_c(block, dest, line_size);
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3356 }
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3357 static void ff_jref_idct_add(uint8_t *dest, int line_size, DCTELEM *block)
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3358 {
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3359 j_rev_dct (block);
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3360 add_pixels_clamped_c(block, dest, line_size);
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3361 }
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3362
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3363 static void ff_jref_idct4_put(uint8_t *dest, int line_size, DCTELEM *block)
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3364 {
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3365 j_rev_dct4 (block);
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3366 put_pixels_clamped4_c(block, dest, line_size);
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3367 }
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3368 static void ff_jref_idct4_add(uint8_t *dest, int line_size, DCTELEM *block)
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3369 {
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3370 j_rev_dct4 (block);
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3371 add_pixels_clamped4_c(block, dest, line_size);
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3372 }
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3373
2257
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3374 static void ff_jref_idct2_put(uint8_t *dest, int line_size, DCTELEM *block)
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3375 {
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3376 j_rev_dct2 (block);
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3377 put_pixels_clamped2_c(block, dest, line_size);
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3378 }
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3379 static void ff_jref_idct2_add(uint8_t *dest, int line_size, DCTELEM *block)
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3380 {
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3381 j_rev_dct2 (block);
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3382 add_pixels_clamped2_c(block, dest, line_size);
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3383 }
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3384
2259
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3385 static void ff_jref_idct1_put(uint8_t *dest, int line_size, DCTELEM *block)
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3386 {
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3387 uint8_t *cm = cropTbl + MAX_NEG_CROP;
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3388
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3389 dest[0] = cm[(block[0] + 4)>>3];
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3390 }
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3391 static void ff_jref_idct1_add(uint8_t *dest, int line_size, DCTELEM *block)
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3392 {
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3393 uint8_t *cm = cropTbl + MAX_NEG_CROP;
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3394
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3395 dest[0] = cm[dest[0] + ((block[0] + 4)>>3)];
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3396 }
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3397
1201
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3398 /* init static data */
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3399 void dsputil_static_init(void)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
3400 {
751
cbe316f082bc warning fixes
michaelni
parents: 706
diff changeset
3401 int i;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
3402
1201
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3403 for(i=0;i<256;i++) cropTbl[i + MAX_NEG_CROP] = i;
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3404 for(i=0;i<MAX_NEG_CROP;i++) {
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3405 cropTbl[i] = 0;
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3406 cropTbl[i + MAX_NEG_CROP + 256] = 255;
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3407 }
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3408
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3409 for(i=0;i<512;i++) {
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3410 squareTbl[i] = (i - 256) * (i - 256);
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3411 }
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3412
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3413 for(i=0; i<64; i++) inv_zigzag_direct16[ff_zigzag_direct[i]]= i+1;
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3414 }
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
3415
861
243cc33da3eb * init for inv_zigzag_direct16 moved to init block
kabi
parents: 860
diff changeset
3416
1201
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3417 void dsputil_init(DSPContext* c, AVCodecContext *avctx)
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3418 {
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3419 int i;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
3420
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3421 #ifdef CONFIG_ENCODERS
1567
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
3422 if(avctx->dct_algo==FF_DCT_FASTINT) {
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3423 c->fdct = fdct_ifast;
1571
aa4dc16c0f18 * adding integer/floating point AAN implementations for DCT 2-4-8
romansh
parents: 1567
diff changeset
3424 c->fdct248 = fdct_ifast248;
1567
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
3425 }
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
3426 else if(avctx->dct_algo==FF_DCT_FAAN) {
1557
5d53c03186a1 floating point AAN DCT
michael
parents: 1527
diff changeset
3427 c->fdct = ff_faandct;
1571
aa4dc16c0f18 * adding integer/floating point AAN implementations for DCT 2-4-8
romansh
parents: 1567
diff changeset
3428 c->fdct248 = ff_faandct248;
1567
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
3429 }
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
3430 else {
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3431 c->fdct = ff_jpeg_fdct_islow; //slow/accurate/default
1567
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
3432 c->fdct248 = ff_fdct248_islow;
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
3433 }
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3434 #endif //CONFIG_ENCODERS
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3435
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3436 if(avctx->lowres==1){
2272
cd43603c46f9 move h264 idct to its own file and call via function pointer in DspContext
michael
parents: 2259
diff changeset
3437 if(avctx->idct_algo==FF_IDCT_INT || avctx->idct_algo==FF_IDCT_AUTO){
cd43603c46f9 move h264 idct to its own file and call via function pointer in DspContext
michael
parents: 2259
diff changeset
3438 c->idct_put= ff_jref_idct4_put;
cd43603c46f9 move h264 idct to its own file and call via function pointer in DspContext
michael
parents: 2259
diff changeset
3439 c->idct_add= ff_jref_idct4_add;
cd43603c46f9 move h264 idct to its own file and call via function pointer in DspContext
michael
parents: 2259
diff changeset
3440 }else{
cd43603c46f9 move h264 idct to its own file and call via function pointer in DspContext
michael
parents: 2259
diff changeset
3441 c->idct_put= ff_h264_lowres_idct_put_c;
cd43603c46f9 move h264 idct to its own file and call via function pointer in DspContext
michael
parents: 2259
diff changeset
3442 c->idct_add= ff_h264_lowres_idct_add_c;
cd43603c46f9 move h264 idct to its own file and call via function pointer in DspContext
michael
parents: 2259
diff changeset
3443 }
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3444 c->idct = j_rev_dct4;
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3445 c->idct_permutation_type= FF_NO_IDCT_PERM;
2257
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3446 }else if(avctx->lowres==2){
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3447 c->idct_put= ff_jref_idct2_put;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3448 c->idct_add= ff_jref_idct2_add;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3449 c->idct = j_rev_dct2;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3450 c->idct_permutation_type= FF_NO_IDCT_PERM;
2259
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3451 }else if(avctx->lowres==3){
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3452 c->idct_put= ff_jref_idct1_put;
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3453 c->idct_add= ff_jref_idct1_add;
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3454 c->idct = j_rev_dct1;
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3455 c->idct_permutation_type= FF_NO_IDCT_PERM;
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3456 }else{
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3457 if(avctx->idct_algo==FF_IDCT_INT){
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3458 c->idct_put= ff_jref_idct_put;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3459 c->idct_add= ff_jref_idct_add;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3460 c->idct = j_rev_dct;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3461 c->idct_permutation_type= FF_LIBMPEG2_IDCT_PERM;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3462 }else{ //accurate/default
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3463 c->idct_put= simple_idct_put;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3464 c->idct_add= simple_idct_add;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3465 c->idct = simple_idct;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3466 c->idct_permutation_type= FF_NO_IDCT_PERM;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3467 }
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3468 }
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3469
2272
cd43603c46f9 move h264 idct to its own file and call via function pointer in DspContext
michael
parents: 2259
diff changeset
3470 c->h264_idct_add= ff_h264_idct_add_c;
cd43603c46f9 move h264 idct to its own file and call via function pointer in DspContext
michael
parents: 2259
diff changeset
3471
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents: 1864
diff changeset
3472 /* VP3 DSP support */
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents: 1864
diff changeset
3473 c->vp3_dsp_init = vp3_dsp_init_c;
1977
89422281f6f6 reorganize and simplify the VP3 IDCT stuff
melanson
parents: 1959
diff changeset
3474 c->vp3_idct = vp3_idct_c;
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents: 1864
diff changeset
3475
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
3476 c->get_pixels = get_pixels_c;
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
3477 c->diff_pixels = diff_pixels_c;
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
3478 c->put_pixels_clamped = put_pixels_clamped_c;
1984
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
3479 c->put_signed_pixels_clamped = put_signed_pixels_clamped_c;
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
3480 c->add_pixels_clamped = add_pixels_clamped_c;
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
3481 c->gmc1 = gmc1_c;
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
3482 c->gmc = gmc_c;
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
3483 c->clear_blocks = clear_blocks_c;
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
3484 c->pix_sum = pix_sum_c;
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
3485 c->pix_norm1 = pix_norm1_c;
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
3486
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3487 /* TODO [0] 16 [1] 8 */
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3488 c->pix_abs[0][0] = pix_abs16_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3489 c->pix_abs[0][1] = pix_abs16_x2_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3490 c->pix_abs[0][2] = pix_abs16_y2_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3491 c->pix_abs[0][3] = pix_abs16_xy2_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3492 c->pix_abs[1][0] = pix_abs8_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3493 c->pix_abs[1][1] = pix_abs8_x2_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3494 c->pix_abs[1][2] = pix_abs8_y2_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3495 c->pix_abs[1][3] = pix_abs8_xy2_c;
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
3496
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3497 #define dspfunc(PFX, IDX, NUM) \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3498 c->PFX ## _pixels_tab[IDX][0] = PFX ## _pixels ## NUM ## _c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3499 c->PFX ## _pixels_tab[IDX][1] = PFX ## _pixels ## NUM ## _x2_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3500 c->PFX ## _pixels_tab[IDX][2] = PFX ## _pixels ## NUM ## _y2_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3501 c->PFX ## _pixels_tab[IDX][3] = PFX ## _pixels ## NUM ## _xy2_c
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
3502
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3503 dspfunc(put, 0, 16);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3504 dspfunc(put_no_rnd, 0, 16);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3505 dspfunc(put, 1, 8);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3506 dspfunc(put_no_rnd, 1, 8);
1267
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
3507 dspfunc(put, 2, 4);
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
3508 dspfunc(put, 3, 2);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
3509
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3510 dspfunc(avg, 0, 16);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3511 dspfunc(avg_no_rnd, 0, 16);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3512 dspfunc(avg, 1, 8);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3513 dspfunc(avg_no_rnd, 1, 8);
1319
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
3514 dspfunc(avg, 2, 4);
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
3515 dspfunc(avg, 3, 2);
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3516 #undef dspfunc
857
b510a7b6decd 100l for kabi (fixing qpel)
michaelni
parents: 853
diff changeset
3517
1864
9a73e6dc5d68 fix motion compensation with (x+1/2,y+1/2) MVs
michael
parents: 1784
diff changeset
3518 c->put_no_rnd_pixels_l2[0]= put_no_rnd_pixels16_l2_c;
9a73e6dc5d68 fix motion compensation with (x+1/2,y+1/2) MVs
michael
parents: 1784
diff changeset
3519 c->put_no_rnd_pixels_l2[1]= put_no_rnd_pixels8_l2_c;
9a73e6dc5d68 fix motion compensation with (x+1/2,y+1/2) MVs
michael
parents: 1784
diff changeset
3520
1267
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
3521 c->put_tpel_pixels_tab[ 0] = put_tpel_pixels_mc00_c;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
3522 c->put_tpel_pixels_tab[ 1] = put_tpel_pixels_mc10_c;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
3523 c->put_tpel_pixels_tab[ 2] = put_tpel_pixels_mc20_c;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
3524 c->put_tpel_pixels_tab[ 4] = put_tpel_pixels_mc01_c;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
3525 c->put_tpel_pixels_tab[ 5] = put_tpel_pixels_mc11_c;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
3526 c->put_tpel_pixels_tab[ 6] = put_tpel_pixels_mc21_c;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
3527 c->put_tpel_pixels_tab[ 8] = put_tpel_pixels_mc02_c;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
3528 c->put_tpel_pixels_tab[ 9] = put_tpel_pixels_mc12_c;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
3529 c->put_tpel_pixels_tab[10] = put_tpel_pixels_mc22_c;
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
3530
1319
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
3531 c->avg_tpel_pixels_tab[ 0] = avg_tpel_pixels_mc00_c;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
3532 c->avg_tpel_pixels_tab[ 1] = avg_tpel_pixels_mc10_c;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
3533 c->avg_tpel_pixels_tab[ 2] = avg_tpel_pixels_mc20_c;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
3534 c->avg_tpel_pixels_tab[ 4] = avg_tpel_pixels_mc01_c;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
3535 c->avg_tpel_pixels_tab[ 5] = avg_tpel_pixels_mc11_c;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
3536 c->avg_tpel_pixels_tab[ 6] = avg_tpel_pixels_mc21_c;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
3537 c->avg_tpel_pixels_tab[ 8] = avg_tpel_pixels_mc02_c;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
3538 c->avg_tpel_pixels_tab[ 9] = avg_tpel_pixels_mc12_c;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
3539 c->avg_tpel_pixels_tab[10] = avg_tpel_pixels_mc22_c;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
3540
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3541 #define dspfunc(PFX, IDX, NUM) \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3542 c->PFX ## _pixels_tab[IDX][ 0] = PFX ## NUM ## _mc00_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3543 c->PFX ## _pixels_tab[IDX][ 1] = PFX ## NUM ## _mc10_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3544 c->PFX ## _pixels_tab[IDX][ 2] = PFX ## NUM ## _mc20_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3545 c->PFX ## _pixels_tab[IDX][ 3] = PFX ## NUM ## _mc30_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3546 c->PFX ## _pixels_tab[IDX][ 4] = PFX ## NUM ## _mc01_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3547 c->PFX ## _pixels_tab[IDX][ 5] = PFX ## NUM ## _mc11_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3548 c->PFX ## _pixels_tab[IDX][ 6] = PFX ## NUM ## _mc21_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3549 c->PFX ## _pixels_tab[IDX][ 7] = PFX ## NUM ## _mc31_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3550 c->PFX ## _pixels_tab[IDX][ 8] = PFX ## NUM ## _mc02_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3551 c->PFX ## _pixels_tab[IDX][ 9] = PFX ## NUM ## _mc12_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3552 c->PFX ## _pixels_tab[IDX][10] = PFX ## NUM ## _mc22_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3553 c->PFX ## _pixels_tab[IDX][11] = PFX ## NUM ## _mc32_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3554 c->PFX ## _pixels_tab[IDX][12] = PFX ## NUM ## _mc03_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3555 c->PFX ## _pixels_tab[IDX][13] = PFX ## NUM ## _mc13_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3556 c->PFX ## _pixels_tab[IDX][14] = PFX ## NUM ## _mc23_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3557 c->PFX ## _pixels_tab[IDX][15] = PFX ## NUM ## _mc33_c
857
b510a7b6decd 100l for kabi (fixing qpel)
michaelni
parents: 853
diff changeset
3558
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3559 dspfunc(put_qpel, 0, 16);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3560 dspfunc(put_no_rnd_qpel, 0, 16);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3561
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3562 dspfunc(avg_qpel, 0, 16);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3563 /* dspfunc(avg_no_rnd_qpel, 0, 16); */
857
b510a7b6decd 100l for kabi (fixing qpel)
michaelni
parents: 853
diff changeset
3564
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3565 dspfunc(put_qpel, 1, 8);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3566 dspfunc(put_no_rnd_qpel, 1, 8);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3567
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3568 dspfunc(avg_qpel, 1, 8);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3569 /* dspfunc(avg_no_rnd_qpel, 1, 8); */
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
3570
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
3571 dspfunc(put_h264_qpel, 0, 16);
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
3572 dspfunc(put_h264_qpel, 1, 8);
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
3573 dspfunc(put_h264_qpel, 2, 4);
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
3574 dspfunc(avg_h264_qpel, 0, 16);
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
3575 dspfunc(avg_h264_qpel, 1, 8);
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
3576 dspfunc(avg_h264_qpel, 2, 4);
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
3577
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3578 #undef dspfunc
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
3579 c->put_h264_chroma_pixels_tab[0]= put_h264_chroma_mc8_c;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
3580 c->put_h264_chroma_pixels_tab[1]= put_h264_chroma_mc4_c;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
3581 c->put_h264_chroma_pixels_tab[2]= put_h264_chroma_mc2_c;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
3582 c->avg_h264_chroma_pixels_tab[0]= avg_h264_chroma_mc8_c;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
3583 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_c;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
3584 c->avg_h264_chroma_pixels_tab[2]= avg_h264_chroma_mc2_c;
857
b510a7b6decd 100l for kabi (fixing qpel)
michaelni
parents: 853
diff changeset
3585
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3586 c->put_mspel_pixels_tab[0]= put_mspel8_mc00_c;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3587 c->put_mspel_pixels_tab[1]= put_mspel8_mc10_c;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3588 c->put_mspel_pixels_tab[2]= put_mspel8_mc20_c;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3589 c->put_mspel_pixels_tab[3]= put_mspel8_mc30_c;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3590 c->put_mspel_pixels_tab[4]= put_mspel8_mc02_c;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3591 c->put_mspel_pixels_tab[5]= put_mspel8_mc12_c;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3592 c->put_mspel_pixels_tab[6]= put_mspel8_mc22_c;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3593 c->put_mspel_pixels_tab[7]= put_mspel8_mc32_c;
1267
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
3594
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3595 #define SET_CMP_FUNC(name) \
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3596 c->name[0]= name ## 16_c;\
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3597 c->name[1]= name ## 8x8_c;
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3598
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3599 SET_CMP_FUNC(hadamard8_diff)
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3600 c->hadamard8_diff[4]= hadamard8_intra16_c;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3601 SET_CMP_FUNC(dct_sad)
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3602 c->sad[0]= pix_abs16_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3603 c->sad[1]= pix_abs8_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3604 c->sse[0]= sse16_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3605 c->sse[1]= sse8_c;
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
3606 c->sse[2]= sse4_c;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3607 SET_CMP_FUNC(quant_psnr)
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3608 SET_CMP_FUNC(rd)
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3609 SET_CMP_FUNC(bit)
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3610 c->vsad[0]= vsad16_c;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3611 c->vsad[4]= vsad_intra16_c;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3612 c->vsse[0]= vsse16_c;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3613 c->vsse[4]= vsse_intra16_c;
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3614 c->nsse[0]= nsse16_c;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3615 c->nsse[1]= nsse8_c;
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
3616 c->w53[0]= w53_16_c;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
3617 c->w53[1]= w53_8_c;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
3618 c->w97[0]= w97_16_c;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
3619 c->w97[1]= w97_8_c;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
3620
866
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3621 c->add_bytes= add_bytes_c;
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3622 c->diff_bytes= diff_bytes_c;
1527
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3623 c->sub_hfyu_median_prediction= sub_hfyu_median_prediction_c;
1273
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
3624 c->bswap_buf= bswap_buf;
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
3625
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
3626 c->h263_h_loop_filter= h263_h_loop_filter_c;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
3627 c->h263_v_loop_filter= h263_v_loop_filter_c;
1784
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3628
2045
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
3629 c->h261_loop_filter= h261_loop_filter_c;
2044
b6f2add2511e h261 decoder by (Maarten Daniels <maarten.daniels at student dot luc dot ac dot be>)
michael
parents: 1984
diff changeset
3630
1784
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3631 c->try_8x8basis= try_8x8basis_c;
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3632 c->add_8x8basis= add_8x8basis_c;
866
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3633
2
2e2c46c87460 fixed config for direct mplayer build compatibility
glantau
parents: 0
diff changeset
3634 #ifdef HAVE_MMX
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3635 dsputil_init_mmx(c, avctx);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
3636 #endif
62
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
3637 #ifdef ARCH_ARMV4L
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3638 dsputil_init_armv4l(c, avctx);
62
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
3639 #endif
88
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
3640 #ifdef HAVE_MLIB
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3641 dsputil_init_mlib(c, avctx);
88
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
3642 #endif
1959
55b7435c59b8 VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents: 1866
diff changeset
3643 #ifdef ARCH_SPARC
55b7435c59b8 VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents: 1866
diff changeset
3644 dsputil_init_vis(c,avctx);
55b7435c59b8 VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents: 1866
diff changeset
3645 #endif
214
73df666cacc7 Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents: 209
diff changeset
3646 #ifdef ARCH_ALPHA
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3647 dsputil_init_alpha(c, avctx);
214
73df666cacc7 Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents: 209
diff changeset
3648 #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
3649 #ifdef ARCH_POWERPC
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3650 dsputil_init_ppc(c, avctx);
626
23a093d6e450 patch by Heliodoro Tammaro <helio at interactives dot org>
michaelni
parents: 625
diff changeset
3651 #endif
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 676
diff changeset
3652 #ifdef HAVE_MMI
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3653 dsputil_init_mmi(c, avctx);
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 676
diff changeset
3654 #endif
1259
e8c3884f2c7e sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents: 1201
diff changeset
3655 #ifdef ARCH_SH4
e8c3884f2c7e sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents: 1201
diff changeset
3656 dsputil_init_sh4(c,avctx);
e8c3884f2c7e sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents: 1201
diff changeset
3657 #endif
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3658
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3659 switch(c->idct_permutation_type){
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3660 case FF_NO_IDCT_PERM:
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3661 for(i=0; i<64; i++)
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3662 c->idct_permutation[i]= i;
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3663 break;
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3664 case FF_LIBMPEG2_IDCT_PERM:
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3665 for(i=0; i<64; i++)
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3666 c->idct_permutation[i]= (i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2);
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3667 break;
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3668 case FF_SIMPLE_IDCT_PERM:
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3669 for(i=0; i<64; i++)
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3670 c->idct_permutation[i]= simple_mmx_permutation[i];
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3671 break;
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3672 case FF_TRANSPOSE_IDCT_PERM:
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3673 for(i=0; i<64; i++)
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3674 c->idct_permutation[i]= ((i&7)<<3) | (i>>3);
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3675 break;
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3676 default:
1598
932d306bf1dc av_log() patch by (Michel Bardiaux <mbardiaux at peaktime dot be>)
michael
parents: 1571
diff changeset
3677 av_log(avctx, AV_LOG_ERROR, "Internal error, IDCT permutation not set\n");
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3678 }
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
3679 }
252
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
3680