annotate dsputil.c @ 3664:de842f000384 libavcodec

Replace usage of put_*_vc1_qpel_pixels_tab with put_*_h264_chroma_pixels_tab
author kostya
date Sat, 02 Sep 2006 04:53:49 +0000
parents 8e180a3276fe
children 9d942506b1f2
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
3036
0b546eab515d Update licensing information: The FSF changed postal address.
diego
parents: 3029
diff changeset
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 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 */
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
22
1106
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 */
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
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"
3198
6b9f0c4fbdbe First part of a series of speed-enchancing patches.
gpoirier
parents: 3105
diff changeset
33 #include "snow.h"
676
c3bdb00a98a9 dont store version for bit-exact tests
michaelni
parents: 651
diff changeset
34
2522
e25782262d7d kill warnings patch by (Mns Rullgrd <mru inprovide com>)
michael
parents: 2448
diff changeset
35 /* snow.c */
e25782262d7d kill warnings patch by (Mns Rullgrd <mru inprovide com>)
michael
parents: 2448
diff changeset
36 void ff_spatial_dwt(int *buffer, int width, int height, int stride, int type, int decomposition_count);
e25782262d7d kill warnings patch by (Mns Rullgrd <mru inprovide com>)
michael
parents: 2448
diff changeset
37
3536
545a15c19c91 sse & sse2 implementations of vorbis channel coupling.
lorenm
parents: 3526
diff changeset
38 /* vorbis.c */
545a15c19c91 sse & sse2 implementations of vorbis channel coupling.
lorenm
parents: 3526
diff changeset
39 void vorbis_inverse_coupling(float *mag, float *ang, int blocksize);
545a15c19c91 sse & sse2 implementations of vorbis channel coupling.
lorenm
parents: 3526
diff changeset
40
2169
db8baace74d8 Minor Patch for shared libs on Mac OSX by (Bill May <wmay at cisco dot com>)
michael
parents: 2066
diff changeset
41 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
42 uint32_t squareTbl[512] = {0, };
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
43
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
44 const uint8_t ff_zigzag_direct[64] = {
706
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 689
diff changeset
45 0, 1, 8, 16, 9, 2, 3, 10,
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 689
diff changeset
46 17, 24, 32, 25, 18, 11, 4, 5,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
47 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
48 27, 20, 13, 6, 7, 14, 21, 28,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
49 35, 42, 49, 56, 57, 50, 43, 36,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
50 29, 22, 15, 23, 30, 37, 44, 51,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
51 58, 59, 52, 45, 38, 31, 39, 46,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
52 53, 60, 61, 54, 47, 55, 62, 63
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
53 };
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
54
1567
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
55 /* Specific zigzag scan for 248 idct. NOTE that unlike the
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
56 specification, we interleave the fields */
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
57 const uint8_t ff_zigzag248_direct[64] = {
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
58 0, 8, 1, 9, 16, 24, 2, 10,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
59 17, 25, 32, 40, 48, 56, 33, 41,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
60 18, 26, 3, 11, 4, 12, 19, 27,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
61 34, 42, 49, 57, 50, 58, 35, 43,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
62 20, 28, 5, 13, 6, 14, 21, 29,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
63 36, 44, 51, 59, 52, 60, 37, 45,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
64 22, 30, 7, 15, 23, 31, 38, 46,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
65 53, 61, 54, 62, 39, 47, 55, 63,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
66 };
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
67
220
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
68 /* not permutated inverse zigzag_direct + 1 for MMX quantizer */
3089
072dbc669253 MSVC-compatible __align8/__align16 declaration
diego
parents: 3036
diff changeset
69 DECLARE_ALIGNED_8(uint16_t, inv_zigzag_direct16[64]) = {0, };
220
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
70
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
71 const uint8_t ff_alternate_horizontal_scan[64] = {
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
72 0, 1, 2, 3, 8, 9, 16, 17,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
73 10, 11, 4, 5, 6, 7, 15, 14,
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
74 13, 12, 19, 18, 24, 25, 32, 33,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
75 26, 27, 20, 21, 22, 23, 28, 29,
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
76 30, 31, 34, 35, 40, 41, 48, 49,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
77 42, 43, 36, 37, 38, 39, 44, 45,
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
78 46, 47, 50, 51, 56, 57, 58, 59,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
79 52, 53, 54, 55, 60, 61, 62, 63,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
80 };
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
81
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
82 const uint8_t ff_alternate_vertical_scan[64] = {
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
83 0, 8, 16, 24, 1, 9, 2, 10,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
84 17, 25, 32, 40, 48, 56, 57, 49,
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
85 41, 33, 26, 18, 3, 11, 4, 12,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
86 19, 27, 34, 42, 50, 58, 35, 43,
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
87 51, 59, 20, 28, 5, 13, 6, 14,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
88 21, 29, 36, 44, 52, 60, 37, 45,
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
89 53, 61, 22, 30, 7, 15, 23, 31,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
90 38, 46, 54, 62, 39, 47, 55, 63,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
91 };
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
92
220
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
93 /* 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
94 const uint32_t inverse[256]={
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
95 0, 4294967295U,2147483648U,1431655766, 1073741824, 858993460, 715827883, 613566757,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
96 536870912, 477218589, 429496730, 390451573, 357913942, 330382100, 306783379, 286331154,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
97 268435456, 252645136, 238609295, 226050911, 214748365, 204522253, 195225787, 186737709,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
98 178956971, 171798692, 165191050, 159072863, 153391690, 148102321, 143165577, 138547333,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
99 134217728, 130150525, 126322568, 122713352, 119304648, 116080198, 113025456, 110127367,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
100 107374183, 104755300, 102261127, 99882961, 97612894, 95443718, 93368855, 91382283,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
101 89478486, 87652394, 85899346, 84215046, 82595525, 81037119, 79536432, 78090315,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
102 76695845, 75350304, 74051161, 72796056, 71582789, 70409300, 69273667, 68174085,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
103 67108864, 66076420, 65075263, 64103990, 63161284, 62245903, 61356676, 60492498,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
104 59652324, 58835169, 58040099, 57266231, 56512728, 55778797, 55063684, 54366675,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
105 53687092, 53024288, 52377650, 51746594, 51130564, 50529028, 49941481, 49367441,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
106 48806447, 48258060, 47721859, 47197443, 46684428, 46182445, 45691142, 45210183,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
107 44739243, 44278014, 43826197, 43383509, 42949673, 42524429, 42107523, 41698712,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
108 41297763, 40904451, 40518560, 40139882, 39768216, 39403370, 39045158, 38693400,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
109 38347923, 38008561, 37675152, 37347542, 37025581, 36709123, 36398028, 36092163,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
110 35791395, 35495598, 35204650, 34918434, 34636834, 34359739, 34087043, 33818641,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
111 33554432, 33294321, 33038210, 32786010, 32537632, 32292988, 32051995, 31814573,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
112 31580642, 31350127, 31122952, 30899046, 30678338, 30460761, 30246249, 30034737,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
113 29826162, 29620465, 29417585, 29217465, 29020050, 28825284, 28633116, 28443493,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
114 28256364, 28071682, 27889399, 27709467, 27531842, 27356480, 27183338, 27012373,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
115 26843546, 26676816, 26512144, 26349493, 26188825, 26030105, 25873297, 25718368,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
116 25565282, 25414008, 25264514, 25116768, 24970741, 24826401, 24683721, 24542671,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
117 24403224, 24265352, 24129030, 23994231, 23860930, 23729102, 23598722, 23469767,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
118 23342214, 23216040, 23091223, 22967740, 22845571, 22724695, 22605092, 22486740,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
119 22369622, 22253717, 22139007, 22025474, 21913099, 21801865, 21691755, 21582751,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
120 21474837, 21367997, 21262215, 21157475, 21053762, 20951060, 20849356, 20748635,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
121 20648882, 20550083, 20452226, 20355296, 20259280, 20164166, 20069941, 19976593,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
122 19884108, 19792477, 19701685, 19611723, 19522579, 19434242, 19346700, 19259944,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
123 19173962, 19088744, 19004281, 18920561, 18837576, 18755316, 18673771, 18592933,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
124 18512791, 18433337, 18354562, 18276457, 18199014, 18122225, 18046082, 17970575,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
125 17895698, 17821442, 17747799, 17674763, 17602325, 17530479, 17459217, 17388532,
220
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
126 17318417, 17248865, 17179870, 17111424, 17043522, 16976156, 16909321, 16843010,
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
127 };
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
128
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
129 /* Input permutation for the simple_idct_mmx */
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
130 static const uint8_t simple_mmx_permutation[64]={
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
131 0x00, 0x08, 0x04, 0x09, 0x01, 0x0C, 0x05, 0x0D,
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
132 0x10, 0x18, 0x14, 0x19, 0x11, 0x1C, 0x15, 0x1D,
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
133 0x20, 0x28, 0x24, 0x29, 0x21, 0x2C, 0x25, 0x2D,
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
134 0x12, 0x1A, 0x16, 0x1B, 0x13, 0x1E, 0x17, 0x1F,
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
135 0x02, 0x0A, 0x06, 0x0B, 0x03, 0x0E, 0x07, 0x0F,
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
136 0x30, 0x38, 0x34, 0x39, 0x31, 0x3C, 0x35, 0x3D,
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
137 0x22, 0x2A, 0x26, 0x2B, 0x23, 0x2E, 0x27, 0x2F,
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
138 0x32, 0x3A, 0x36, 0x3B, 0x33, 0x3E, 0x37, 0x3F,
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
139 };
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
140
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
141 static int pix_sum_c(uint8_t * pix, int line_size)
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
142 {
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
143 int s, i, j;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
144
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
145 s = 0;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
146 for (i = 0; i < 16; i++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
147 for (j = 0; j < 16; j += 8) {
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
148 s += pix[0];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
149 s += pix[1];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
150 s += pix[2];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
151 s += pix[3];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
152 s += pix[4];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
153 s += pix[5];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
154 s += pix[6];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
155 s += pix[7];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
156 pix += 8;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
157 }
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
158 pix += line_size - 16;
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
159 }
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
160 return s;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
161 }
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
162
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
163 static int pix_norm1_c(uint8_t * pix, int line_size)
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
164 {
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
165 int s, i, j;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
166 uint32_t *sq = squareTbl + 256;
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
167
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
168 s = 0;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
169 for (i = 0; i < 16; i++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
170 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
171 #if 0
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
172 s += sq[pix[0]];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
173 s += sq[pix[1]];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
174 s += sq[pix[2]];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
175 s += sq[pix[3]];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
176 s += sq[pix[4]];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
177 s += sq[pix[5]];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
178 s += sq[pix[6]];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
179 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
180 #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
181 #if LONG_MAX > 2147483647
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
182 register uint64_t x=*(uint64_t*)pix;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
183 s += sq[x&0xff];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
184 s += sq[(x>>8)&0xff];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
185 s += sq[(x>>16)&0xff];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
186 s += sq[(x>>24)&0xff];
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
187 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
188 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
189 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
190 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
191 #else
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
192 register uint32_t x=*(uint32_t*)pix;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
193 s += sq[x&0xff];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
194 s += sq[(x>>8)&0xff];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
195 s += sq[(x>>16)&0xff];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
196 s += sq[(x>>24)&0xff];
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
197 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
198 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
199 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
200 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
201 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
202 #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
203 #endif
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
204 pix += 8;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
205 }
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
206 pix += line_size - 16;
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
207 }
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
208 return s;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
209 }
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
210
1273
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
211 static void bswap_buf(uint32_t *dst, uint32_t *src, int w){
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
212 int i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
213
1273
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
214 for(i=0; i+8<=w; i+=8){
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
215 dst[i+0]= bswap_32(src[i+0]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
216 dst[i+1]= bswap_32(src[i+1]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
217 dst[i+2]= bswap_32(src[i+2]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
218 dst[i+3]= bswap_32(src[i+3]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
219 dst[i+4]= bswap_32(src[i+4]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
220 dst[i+5]= bswap_32(src[i+5]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
221 dst[i+6]= bswap_32(src[i+6]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
222 dst[i+7]= bswap_32(src[i+7]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
223 }
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
224 for(;i<w; i++){
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
225 dst[i+0]= bswap_32(src[i+0]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
226 }
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
227 }
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
228
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
229 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
230 {
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
231 int s, i;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
232 uint32_t *sq = squareTbl + 256;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
233
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
234 s = 0;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
235 for (i = 0; i < h; i++) {
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
236 s += sq[pix1[0] - pix2[0]];
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
237 s += sq[pix1[1] - pix2[1]];
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
238 s += sq[pix1[2] - pix2[2]];
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
239 s += sq[pix1[3] - pix2[3]];
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
240 pix1 += line_size;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
241 pix2 += line_size;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
242 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
243 return s;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
244 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
245
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
246 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
247 {
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
248 int s, i;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
249 uint32_t *sq = squareTbl + 256;
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
250
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
251 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
252 for (i = 0; i < h; i++) {
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
253 s += sq[pix1[0] - pix2[0]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
254 s += sq[pix1[1] - pix2[1]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
255 s += sq[pix1[2] - pix2[2]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
256 s += sq[pix1[3] - pix2[3]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
257 s += sq[pix1[4] - pix2[4]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
258 s += sq[pix1[5] - pix2[5]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
259 s += sq[pix1[6] - pix2[6]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
260 s += sq[pix1[7] - pix2[7]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
261 pix1 += line_size;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
262 pix2 += line_size;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
263 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
264 return s;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
265 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
266
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
267 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
268 {
1012
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
269 int s, i;
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
270 uint32_t *sq = squareTbl + 256;
884
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 882
diff changeset
271
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 882
diff changeset
272 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
273 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
274 s += sq[pix1[ 0] - pix2[ 0]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
275 s += sq[pix1[ 1] - pix2[ 1]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
276 s += sq[pix1[ 2] - pix2[ 2]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
277 s += sq[pix1[ 3] - pix2[ 3]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
278 s += sq[pix1[ 4] - pix2[ 4]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
279 s += sq[pix1[ 5] - pix2[ 5]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
280 s += sq[pix1[ 6] - pix2[ 6]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
281 s += sq[pix1[ 7] - pix2[ 7]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
282 s += sq[pix1[ 8] - pix2[ 8]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
283 s += sq[pix1[ 9] - pix2[ 9]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
284 s += sq[pix1[10] - pix2[10]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
285 s += sq[pix1[11] - pix2[11]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
286 s += sq[pix1[12] - pix2[12]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
287 s += sq[pix1[13] - pix2[13]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
288 s += sq[pix1[14] - pix2[14]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
289 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
290
1012
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
291 pix1 += line_size;
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
292 pix2 += line_size;
884
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 882
diff changeset
293 }
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 882
diff changeset
294 return s;
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 882
diff changeset
295 }
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 882
diff changeset
296
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
297
3373
b8996cc5ccae Disable w53 and w97 cmp methods when snow encoder is disabled
gpoirier
parents: 3323
diff changeset
298 #ifdef CONFIG_SNOW_ENCODER //dwt is in snow.c
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
299 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
300 int s, i, j;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
301 const int dec_count= w==8 ? 3 : 4;
3323
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
302 int tmp[32*32];
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
303 int level, ori;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
304 static const int scale[2][2][4][4]={
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
305 {
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
306 {
3323
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
307 // 9/7 8x8 dec=3
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
308 {268, 239, 239, 213},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
309 { 0, 224, 224, 152},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
310 { 0, 135, 135, 110},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
311 },{
3323
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
312 // 9/7 16x16 or 32x32 dec=4
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
313 {344, 310, 310, 280},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
314 { 0, 320, 320, 228},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
315 { 0, 175, 175, 136},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
316 { 0, 129, 129, 102},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
317 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
318 },{
3323
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
319 {
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
320 // 5/3 8x8 dec=3
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
321 {275, 245, 245, 218},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
322 { 0, 230, 230, 156},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
323 { 0, 138, 138, 113},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
324 },{
3323
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
325 // 5/3 16x16 or 32x32 dec=4
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
326 {352, 317, 317, 286},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
327 { 0, 328, 328, 233},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
328 { 0, 180, 180, 140},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
329 { 0, 132, 132, 105},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
330 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
331 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
332 };
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
333
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
334 for (i = 0; i < h; i++) {
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
335 for (j = 0; j < w; j+=4) {
3323
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
336 tmp[32*i+j+0] = (pix1[j+0] - pix2[j+0])<<4;
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
337 tmp[32*i+j+1] = (pix1[j+1] - pix2[j+1])<<4;
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
338 tmp[32*i+j+2] = (pix1[j+2] - pix2[j+2])<<4;
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
339 tmp[32*i+j+3] = (pix1[j+3] - pix2[j+3])<<4;
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
340 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
341 pix1 += line_size;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
342 pix2 += line_size;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
343 }
2639
e2780f828440 put most codecs under ifdefs
michael
parents: 2633
diff changeset
344
3323
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
345 ff_spatial_dwt(tmp, w, h, 32, type, dec_count);
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
346
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
347 s=0;
3323
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
348 assert(w==h);
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
349 for(level=0; level<dec_count; level++){
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
350 for(ori= level ? 1 : 0; ori<4; ori++){
3323
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
351 int size= w>>(dec_count-level);
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
352 int sx= (ori&1) ? size : 0;
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
353 int stride= 32<<(dec_count-level);
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
354 int sy= (ori&2) ? stride>>1 : 0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
355
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
356 for(i=0; i<size; i++){
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
357 for(j=0; j<size; j++){
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
358 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
359 s += ABS(v);
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
360 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
361 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
362 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
363 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
364 assert(s>=0);
3323
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
365 return s>>9;
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
366 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
367
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
368 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
369 return w_c(v, pix1, pix2, line_size, 8, h, 1);
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
370 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
371
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
372 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
373 return w_c(v, pix1, pix2, line_size, 8, h, 0);
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
374 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
375
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
376 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
377 return w_c(v, pix1, pix2, line_size, 16, h, 1);
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
378 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
379
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
380 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
381 return w_c(v, pix1, pix2, line_size, 16, h, 0);
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
382 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
383
3323
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
384 int w53_32_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h){
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
385 return w_c(v, pix1, pix2, line_size, 32, h, 1);
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
386 }
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
387
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
388 int w97_32_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h){
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
389 return w_c(v, pix1, pix2, line_size, 32, h, 0);
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
390 }
3373
b8996cc5ccae Disable w53 and w97 cmp methods when snow encoder is disabled
gpoirier
parents: 3323
diff changeset
391 #endif
3323
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
392
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
393 static void get_pixels_c(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
394 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
395 int i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
396
986e461dc072 Initial revision
glantau
parents:
diff changeset
397 /* read the pixels */
986e461dc072 Initial revision
glantau
parents:
diff changeset
398 for(i=0;i<8;i++) {
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
399 block[0] = pixels[0];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
400 block[1] = pixels[1];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
401 block[2] = pixels[2];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
402 block[3] = pixels[3];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
403 block[4] = pixels[4];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
404 block[5] = pixels[5];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
405 block[6] = pixels[6];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
406 block[7] = pixels[7];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
407 pixels += line_size;
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
408 block += 8;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
409 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
410 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
411
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
412 static void diff_pixels_c(DCTELEM *restrict block, const uint8_t *s1,
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
413 const uint8_t *s2, int stride){
324
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
414 int i;
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
415
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
416 /* read the pixels */
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
417 for(i=0;i<8;i++) {
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
418 block[0] = s1[0] - s2[0];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
419 block[1] = s1[1] - s2[1];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
420 block[2] = s1[2] - s2[2];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
421 block[3] = s1[3] - s2[3];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
422 block[4] = s1[4] - s2[4];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
423 block[5] = s1[5] - s2[5];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
424 block[6] = s1[6] - s2[6];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
425 block[7] = s1[7] - s2[7];
324
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
426 s1 += stride;
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
427 s2 += stride;
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
428 block += 8;
324
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
429 }
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
430 }
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
431
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
432
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
433 static void put_pixels_clamped_c(const DCTELEM *block, uint8_t *restrict pixels,
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
434 int line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
435 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
436 int i;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
437 uint8_t *cm = cropTbl + MAX_NEG_CROP;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
438
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
439 /* read the pixels */
986e461dc072 Initial revision
glantau
parents:
diff changeset
440 for(i=0;i<8;i++) {
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
441 pixels[0] = cm[block[0]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
442 pixels[1] = cm[block[1]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
443 pixels[2] = cm[block[2]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
444 pixels[3] = cm[block[3]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
445 pixels[4] = cm[block[4]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
446 pixels[5] = cm[block[5]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
447 pixels[6] = cm[block[6]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
448 pixels[7] = cm[block[7]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
449
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
450 pixels += line_size;
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
451 block += 8;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
452 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
453 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
454
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
455 static void put_pixels_clamped4_c(const DCTELEM *block, uint8_t *restrict pixels,
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
456 int line_size)
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
457 {
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
458 int i;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
459 uint8_t *cm = cropTbl + MAX_NEG_CROP;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
460
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
461 /* read the pixels */
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
462 for(i=0;i<4;i++) {
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
463 pixels[0] = cm[block[0]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
464 pixels[1] = cm[block[1]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
465 pixels[2] = cm[block[2]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
466 pixels[3] = cm[block[3]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
467
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
468 pixels += line_size;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
469 block += 8;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
470 }
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
471 }
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
472
2257
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
473 static void put_pixels_clamped2_c(const DCTELEM *block, uint8_t *restrict pixels,
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
474 int line_size)
2257
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
475 {
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
476 int i;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
477 uint8_t *cm = cropTbl + MAX_NEG_CROP;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
478
2257
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
479 /* read the pixels */
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
480 for(i=0;i<2;i++) {
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
481 pixels[0] = cm[block[0]];
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
482 pixels[1] = cm[block[1]];
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
483
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
484 pixels += line_size;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
485 block += 8;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
486 }
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
487 }
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
488
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
489 static void put_signed_pixels_clamped_c(const DCTELEM *block,
1984
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
490 uint8_t *restrict pixels,
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
491 int line_size)
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
492 {
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
493 int i, j;
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
494
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
495 for (i = 0; i < 8; i++) {
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
496 for (j = 0; j < 8; j++) {
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
497 if (*block < -128)
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
498 *pixels = 0;
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
499 else if (*block > 127)
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
500 *pixels = 255;
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
501 else
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
502 *pixels = (uint8_t)(*block + 128);
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
503 block++;
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
504 pixels++;
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
505 }
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
506 pixels += (line_size - 8);
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
507 }
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
508 }
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
509
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
510 static void add_pixels_clamped_c(const DCTELEM *block, uint8_t *restrict pixels,
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
511 int line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
512 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
513 int i;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
514 uint8_t *cm = cropTbl + MAX_NEG_CROP;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
515
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
516 /* read the pixels */
986e461dc072 Initial revision
glantau
parents:
diff changeset
517 for(i=0;i<8;i++) {
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
518 pixels[0] = cm[pixels[0] + block[0]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
519 pixels[1] = cm[pixels[1] + block[1]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
520 pixels[2] = cm[pixels[2] + block[2]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
521 pixels[3] = cm[pixels[3] + block[3]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
522 pixels[4] = cm[pixels[4] + block[4]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
523 pixels[5] = cm[pixels[5] + block[5]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
524 pixels[6] = cm[pixels[6] + block[6]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
525 pixels[7] = cm[pixels[7] + block[7]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
526 pixels += line_size;
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
527 block += 8;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
528 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
529 }
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
530
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
531 static void add_pixels_clamped4_c(const DCTELEM *block, uint8_t *restrict pixels,
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
532 int line_size)
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
533 {
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
534 int i;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
535 uint8_t *cm = cropTbl + MAX_NEG_CROP;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
536
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
537 /* read the pixels */
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
538 for(i=0;i<4;i++) {
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
539 pixels[0] = cm[pixels[0] + block[0]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
540 pixels[1] = cm[pixels[1] + block[1]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
541 pixels[2] = cm[pixels[2] + block[2]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
542 pixels[3] = cm[pixels[3] + block[3]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
543 pixels += line_size;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
544 block += 8;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
545 }
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
546 }
2257
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
547
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
548 static void add_pixels_clamped2_c(const DCTELEM *block, uint8_t *restrict pixels,
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
549 int line_size)
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
550 {
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
551 int i;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
552 uint8_t *cm = cropTbl + MAX_NEG_CROP;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
553
2257
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
554 /* read the pixels */
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
555 for(i=0;i<2;i++) {
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
556 pixels[0] = cm[pixels[0] + block[0]];
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
557 pixels[1] = cm[pixels[1] + block[1]];
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
558 pixels += line_size;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
559 block += 8;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
560 }
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
561 }
2763
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
562
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
563 static void add_pixels8_c(uint8_t *restrict pixels, DCTELEM *block, int line_size)
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
564 {
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
565 int i;
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
566 for(i=0;i<8;i++) {
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
567 pixels[0] += block[0];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
568 pixels[1] += block[1];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
569 pixels[2] += block[2];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
570 pixels[3] += block[3];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
571 pixels[4] += block[4];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
572 pixels[5] += block[5];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
573 pixels[6] += block[6];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
574 pixels[7] += block[7];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
575 pixels += line_size;
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
576 block += 8;
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
577 }
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
578 }
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
579
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
580 static void add_pixels4_c(uint8_t *restrict pixels, DCTELEM *block, int line_size)
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
581 {
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
582 int i;
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
583 for(i=0;i<4;i++) {
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
584 pixels[0] += block[0];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
585 pixels[1] += block[1];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
586 pixels[2] += block[2];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
587 pixels[3] += block[3];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
588 pixels += line_size;
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
589 block += 4;
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
590 }
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
591 }
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
592
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
593 #if 0
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
594
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
595 #define PIXOP2(OPNAME, OP) \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
596 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
597 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
598 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
599 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
600 OP(*((uint64_t*)block), LD64(pixels));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
601 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
602 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
603 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
604 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
605 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
606 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
607 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
608 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
609 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
610 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
611 const uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
612 OP(*((uint64_t*)block), (a&b) + (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
613 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
614 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
615 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
616 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
617 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
618 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
619 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
620 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
621 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
622 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
623 const uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
624 OP(*((uint64_t*)block), (a|b) - (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
625 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
626 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
627 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
628 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
629 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
630 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
631 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
632 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
633 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
634 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
635 const uint64_t b= LD64(pixels+line_size);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
636 OP(*((uint64_t*)block), (a&b) + (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
637 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
638 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
639 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
640 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
641 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
642 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
643 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
644 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
645 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
646 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
647 const uint64_t b= LD64(pixels+line_size);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
648 OP(*((uint64_t*)block), (a|b) - (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
649 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
650 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
651 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
652 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
653 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
654 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
655 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
656 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
657 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
658 const uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
659 uint64_t l0= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
660 + (b&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
661 + 0x0202020202020202ULL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
662 uint64_t h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
663 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
664 uint64_t l1,h1;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
665 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
666 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
667 for(i=0; i<h; i+=2){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
668 uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
669 uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
670 l1= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
671 + (b&0x0303030303030303ULL);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
672 h1= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
673 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
674 OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
675 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
676 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
677 a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
678 b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
679 l0= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
680 + (b&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
681 + 0x0202020202020202ULL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
682 h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
683 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
684 OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
685 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
686 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
687 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
688 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
689 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
690 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
691 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
692 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
693 const uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
694 const uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
695 uint64_t l0= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
696 + (b&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
697 + 0x0101010101010101ULL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
698 uint64_t h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
699 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
700 uint64_t l1,h1;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
701 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
702 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
703 for(i=0; i<h; i+=2){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
704 uint64_t a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
705 uint64_t b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
706 l1= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
707 + (b&0x0303030303030303ULL);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
708 h1= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
709 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
710 OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
711 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
712 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
713 a= LD64(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
714 b= LD64(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
715 l0= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
716 + (b&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
717 + 0x0101010101010101ULL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
718 h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
719 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
720 OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
721 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
722 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
723 }\
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 CALL_2X_PIXELS(OPNAME ## _pixels16_c , OPNAME ## _pixels_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
727 CALL_2X_PIXELS(OPNAME ## _pixels16_x2_c , OPNAME ## _pixels_x2_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
728 CALL_2X_PIXELS(OPNAME ## _pixels16_y2_c , OPNAME ## _pixels_y2_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
729 CALL_2X_PIXELS(OPNAME ## _pixels16_xy2_c, OPNAME ## _pixels_xy2_c, 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
730 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
731 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
732 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
733
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
734 #define op_avg(a, b) a = ( ((a)|(b)) - ((((a)^(b))&0xFEFEFEFEFEFEFEFEULL)>>1) )
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
735 #else // 64 bit variant
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
736
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
737 #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
738 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
739 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
740 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
741 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
742 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
743 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
744 }\
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
745 }\
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
746 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
747 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
748 for(i=0; i<h; i++){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
749 OP(*((uint32_t*)(block )), LD32(pixels ));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
750 pixels+=line_size;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
751 block +=line_size;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
752 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
753 }\
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
754 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
755 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
756 for(i=0; i<h; i++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
757 OP(*((uint32_t*)(block )), LD32(pixels ));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
758 OP(*((uint32_t*)(block+4)), LD32(pixels+4));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
759 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
760 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
761 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
762 }\
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
763 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
764 OPNAME ## _pixels8_c(block, pixels, line_size, h);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
765 }\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
766 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
767 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
768 int src_stride1, int src_stride2, int h){\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
769 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
770 for(i=0; i<h; i++){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
771 uint32_t a,b;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
772 a= LD32(&src1[i*src_stride1 ]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
773 b= LD32(&src2[i*src_stride2 ]);\
1264
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
774 OP(*((uint32_t*)&dst[i*dst_stride ]), no_rnd_avg32(a, b));\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
775 a= LD32(&src1[i*src_stride1+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
776 b= LD32(&src2[i*src_stride2+4]);\
1264
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
777 OP(*((uint32_t*)&dst[i*dst_stride+4]), no_rnd_avg32(a, b));\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
778 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
779 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
780 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
781 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
782 int src_stride1, int src_stride2, int h){\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
783 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
784 for(i=0; i<h; i++){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
785 uint32_t a,b;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
786 a= LD32(&src1[i*src_stride1 ]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
787 b= LD32(&src2[i*src_stride2 ]);\
1264
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
788 OP(*((uint32_t*)&dst[i*dst_stride ]), rnd_avg32(a, b));\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
789 a= LD32(&src1[i*src_stride1+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
790 b= LD32(&src2[i*src_stride2+4]);\
1264
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
791 OP(*((uint32_t*)&dst[i*dst_stride+4]), rnd_avg32(a, b));\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
792 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
793 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
794 \
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
795 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
796 int src_stride1, int src_stride2, int h){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
797 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
798 for(i=0; i<h; i++){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
799 uint32_t a,b;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
800 a= LD32(&src1[i*src_stride1 ]);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
801 b= LD32(&src2[i*src_stride2 ]);\
1264
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
802 OP(*((uint32_t*)&dst[i*dst_stride ]), rnd_avg32(a, b));\
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
803 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
804 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
805 \
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
806 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
807 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
808 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
809 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
810 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
811 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
812 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
813 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
814 }\
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
815 }\
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
816 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
817 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
818 int src_stride1, int src_stride2, int h){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
819 OPNAME ## _pixels8_l2(dst , src1 , src2 , dst_stride, src_stride1, src_stride2, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
820 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
821 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
822 \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
823 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
824 int src_stride1, int src_stride2, int h){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
825 OPNAME ## _no_rnd_pixels8_l2(dst , src1 , src2 , dst_stride, src_stride1, src_stride2, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
826 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
827 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
828 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
829 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
830 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
831 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
832 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
833 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
834 OPNAME ## _pixels8_l2(block, pixels, pixels+1, line_size, line_size, line_size, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
835 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
836 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
837 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
838 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
839 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
840 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
841 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
842 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
843 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
844 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
845 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
846 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
847 int i;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
848 for(i=0; i<h; i++){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
849 uint32_t a, b, c, d, l0, l1, h0, h1;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
850 a= LD32(&src1[i*src_stride1]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
851 b= LD32(&src2[i*src_stride2]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
852 c= LD32(&src3[i*src_stride3]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
853 d= LD32(&src4[i*src_stride4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
854 l0= (a&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
855 + (b&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
856 + 0x02020202UL;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
857 h0= ((a&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
858 + ((b&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
859 l1= (c&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
860 + (d&0x03030303UL);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
861 h1= ((c&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
862 + ((d&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
863 OP(*((uint32_t*)&dst[i*dst_stride]), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
864 a= LD32(&src1[i*src_stride1+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
865 b= LD32(&src2[i*src_stride2+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
866 c= LD32(&src3[i*src_stride3+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
867 d= LD32(&src4[i*src_stride4+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
868 l0= (a&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
869 + (b&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
870 + 0x02020202UL;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
871 h0= ((a&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
872 + ((b&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
873 l1= (c&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
874 + (d&0x03030303UL);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
875 h1= ((c&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
876 + ((d&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
877 OP(*((uint32_t*)&dst[i*dst_stride+4]), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
878 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
879 }\
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
880 \
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
881 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
882 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
883 }\
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
884 \
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
885 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
886 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
887 }\
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
888 \
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
889 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
890 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
891 }\
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
892 \
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
893 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
894 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
895 }\
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
896 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
897 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
898 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
899 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
900 for(i=0; i<h; i++){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
901 uint32_t a, b, c, d, l0, l1, h0, h1;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
902 a= LD32(&src1[i*src_stride1]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
903 b= LD32(&src2[i*src_stride2]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
904 c= LD32(&src3[i*src_stride3]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
905 d= LD32(&src4[i*src_stride4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
906 l0= (a&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
907 + (b&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
908 + 0x01010101UL;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
909 h0= ((a&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
910 + ((b&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
911 l1= (c&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
912 + (d&0x03030303UL);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
913 h1= ((c&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
914 + ((d&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
915 OP(*((uint32_t*)&dst[i*dst_stride]), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
916 a= LD32(&src1[i*src_stride1+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
917 b= LD32(&src2[i*src_stride2+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
918 c= LD32(&src3[i*src_stride3+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
919 d= LD32(&src4[i*src_stride4+4]);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
920 l0= (a&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
921 + (b&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
922 + 0x01010101UL;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
923 h0= ((a&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
924 + ((b&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
925 l1= (c&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
926 + (d&0x03030303UL);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
927 h1= ((c&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
928 + ((d&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
929 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
930 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
931 }\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
932 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
933 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
934 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
935 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
936 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
937 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
938 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
939 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
940 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
941 }\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
942 \
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
943 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
944 {\
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 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
946 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
947 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
948 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
949 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
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 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
954 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
955 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
956 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
957 \
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 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
959 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
960 \
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 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
962 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
963 \
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 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
965 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
966 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
967 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
968 \
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 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
970 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
971 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
972 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
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 }\
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
975 \
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
976 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
977 {\
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
978 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
979 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
980 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
981 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
982 + (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
983 + 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
984 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
985 + ((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
986 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
987 \
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
988 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
989 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
990 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
991 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
992 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
993 + (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
994 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
995 + ((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
996 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
997 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
998 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
999 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
1000 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
1001 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
1002 + (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
1003 + 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
1004 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
1005 + ((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
1006 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
1007 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
1008 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
1009 }\
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
1010 }\
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
1011 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1012 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
1013 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1014 int j;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1015 for(j=0; j<2; j++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1016 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1017 const uint32_t a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1018 const uint32_t b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1019 uint32_t l0= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1020 + (b&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1021 + 0x02020202UL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1022 uint32_t h0= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1023 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1024 uint32_t l1,h1;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1025 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1026 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1027 for(i=0; i<h; i+=2){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1028 uint32_t a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1029 uint32_t b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1030 l1= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1031 + (b&0x03030303UL);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1032 h1= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1033 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1034 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1035 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1036 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1037 a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1038 b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1039 l0= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1040 + (b&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1041 + 0x02020202UL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1042 h0= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1043 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1044 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1045 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1046 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1047 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1048 pixels+=4-line_size*(h+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1049 block +=4-line_size*h;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1050 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1051 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1052 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1053 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
1054 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1055 int j;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1056 for(j=0; j<2; j++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1057 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1058 const uint32_t a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1059 const uint32_t b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1060 uint32_t l0= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1061 + (b&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1062 + 0x01010101UL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1063 uint32_t h0= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1064 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1065 uint32_t l1,h1;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1066 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1067 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1068 for(i=0; i<h; i+=2){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1069 uint32_t a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1070 uint32_t b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1071 l1= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1072 + (b&0x03030303UL);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1073 h1= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1074 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1075 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1076 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1077 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1078 a= LD32(pixels );\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1079 b= LD32(pixels+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1080 l0= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1081 + (b&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1082 + 0x01010101UL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1083 h0= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1084 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1085 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1086 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1087 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1088 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1089 pixels+=4-line_size*(h+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1090 block +=4-line_size*h;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1091 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1092 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1093 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1094 CALL_2X_PIXELS(OPNAME ## _pixels16_c , OPNAME ## _pixels8_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1095 CALL_2X_PIXELS(OPNAME ## _pixels16_x2_c , OPNAME ## _pixels8_x2_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1096 CALL_2X_PIXELS(OPNAME ## _pixels16_y2_c , OPNAME ## _pixels8_y2_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1097 CALL_2X_PIXELS(OPNAME ## _pixels16_xy2_c, OPNAME ## _pixels8_xy2_c, 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1098 CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_c , OPNAME ## _pixels8_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1099 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
1100 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
1101 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
1102
1264
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
1103 #define op_avg(a, b) a = rnd_avg32(a, b)
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1104 #endif
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1105 #define op_put(a, b) a = b
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1106
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1107 PIXOP2(avg, op_avg)
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1108 PIXOP2(put, op_put)
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1109 #undef op_avg
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1110 #undef op_put
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1111
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1112 #define avg2(a,b) ((a+b+1)>>1)
986e461dc072 Initial revision
glantau
parents:
diff changeset
1113 #define avg4(a,b,c,d) ((a+b+c+d+2)>>2)
986e461dc072 Initial revision
glantau
parents:
diff changeset
1114
1864
9a73e6dc5d68 fix motion compensation with (x+1/2,y+1/2) MVs
michael
parents: 1784
diff changeset
1115 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
1116 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
1117 }
9a73e6dc5d68 fix motion compensation with (x+1/2,y+1/2) MVs
michael
parents: 1784
diff changeset
1118
9a73e6dc5d68 fix motion compensation with (x+1/2,y+1/2) MVs
michael
parents: 1784
diff changeset
1119 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
1120 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
1121 }
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1122
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1123 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
1124 {
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1125 const int A=(16-x16)*(16-y16);
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1126 const int B=( x16)*(16-y16);
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1127 const int C=(16-x16)*( y16);
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1128 const int D=( x16)*( y16);
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1129 int i;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1130
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1131 for(i=0; i<h; i++)
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1132 {
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1133 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
1134 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
1135 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
1136 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
1137 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
1138 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
1139 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
1140 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
1141 dst+= stride;
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1142 src+= stride;
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1143 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1144 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1145
3248
7aa9f80e7954 mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents: 3245
diff changeset
1146 void ff_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
1147 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
1148 {
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1149 int y, vx, vy;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1150 const int s= 1<<shift;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
1151
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1152 width--;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1153 height--;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1154
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1155 for(y=0; y<h; y++){
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1156 int x;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1157
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1158 vx= ox;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1159 vy= oy;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1160 for(x=0; x<8; x++){ //XXX FIXME optimize
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1161 int src_x, src_y, frac_x, frac_y, index;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1162
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1163 src_x= vx>>16;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1164 src_y= vy>>16;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1165 frac_x= src_x&(s-1);
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1166 frac_y= src_y&(s-1);
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1167 src_x>>=shift;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1168 src_y>>=shift;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
1169
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1170 if((unsigned)src_x < width){
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1171 if((unsigned)src_y < height){
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1172 index= src_x + src_y*stride;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1173 dst[y*stride + x]= ( ( src[index ]*(s-frac_x)
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1174 + src[index +1]* frac_x )*(s-frac_y)
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1175 + ( src[index+stride ]*(s-frac_x)
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1176 + src[index+stride+1]* frac_x )* frac_y
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1177 + r)>>(shift*2);
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1178 }else{
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
1179 index= src_x + clip(src_y, 0, height)*stride;
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
1180 dst[y*stride + x]= ( ( src[index ]*(s-frac_x)
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1181 + src[index +1]* frac_x )*s
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1182 + r)>>(shift*2);
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1183 }
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1184 }else{
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1185 if((unsigned)src_y < height){
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
1186 index= clip(src_x, 0, width) + src_y*stride;
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
1187 dst[y*stride + x]= ( ( src[index ]*(s-frac_y)
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1188 + src[index+stride ]* frac_y )*s
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1189 + r)>>(shift*2);
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1190 }else{
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
1191 index= clip(src_x, 0, width) + clip(src_y, 0, height)*stride;
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1192 dst[y*stride + x]= src[index ];
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1193 }
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1194 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
1195
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1196 vx+= dxx;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1197 vy+= dyx;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1198 }
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1199 ox += dxy;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1200 oy += dyy;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1201 }
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1202 }
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
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 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
1205 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
1206 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
1207 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
1208 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
1209 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
1210 }
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 }
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
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 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
1214 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
1215 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
1216 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1217 dst[j] = (683*(2*src[j] + src[j+1] + 1)) >> 11;
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
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 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
1220 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
1221 }
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 }
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
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
1224 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
1225 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
1226 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
1227 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1228 dst[j] = (683*(src[j] + 2*src[j+1] + 1)) >> 11;
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
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 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
1231 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
1232 }
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 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
1234
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
1235 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
1236 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
1237 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
1238 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1239 dst[j] = (683*(2*src[j] + src[j+stride] + 1)) >> 11;
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
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 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
1242 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
1243 }
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 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
1245
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
1246 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
1247 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
1248 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
1249 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1250 dst[j] = (2731*(4*src[j] + 3*src[j+1] + 3*src[j+stride] + 2*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
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 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
1253 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
1254 }
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 }
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
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 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
1258 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
1259 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
1260 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1261 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
1262 }
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
1263 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
1264 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
1265 }
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
1266 }
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
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
1268 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
1269 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
1270 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
1271 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1272 dst[j] = (683*(src[j] + 2*src[j+stride] + 1)) >> 11;
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
1273 }
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
1274 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
1275 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
1276 }
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
1277 }
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
1278
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
1279 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
1280 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
1281 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
1282 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1283 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
1284 }
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
1285 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
1286 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
1287 }
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
1288 }
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
1289
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
1290 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
1291 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
1292 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
1293 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1294 dst[j] = (2731*(2*src[j] + 3*src[j+1] + 3*src[j+stride] + 4*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
1295 }
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
1296 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
1297 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
1298 }
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
1299 }
1319
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 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
1302 switch(width){
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1303 case 2: avg_pixels2_c (dst, src, stride, height); break;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1304 case 4: avg_pixels4_c (dst, src, stride, height); break;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1305 case 8: avg_pixels8_c (dst, src, stride, height); break;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1306 case 16:avg_pixels16_c(dst, src, stride, height); break;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1307 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1308 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1309
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1310 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
1311 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1312 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1313 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1314 dst[j] = (dst[j] + ((683*(2*src[j] + src[j+1] + 1)) >> 11) + 1) >> 1;
1319
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 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1317 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1318 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1319 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1320
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1321 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
1322 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1323 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1324 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1325 dst[j] = (dst[j] + ((683*(src[j] + 2*src[j+1] + 1)) >> 11) + 1) >> 1;
1319
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 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1328 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1329 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1330 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
1331
1319
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1332 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
1333 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1334 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1335 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1336 dst[j] = (dst[j] + ((683*(2*src[j] + src[j+stride] + 1)) >> 11) + 1) >> 1;
1319
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 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1339 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1340 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1341 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
1342
1319
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1343 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
1344 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1345 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1346 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1347 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;
1319
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 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1350 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1351 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1352 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1353
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1354 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
1355 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1356 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1357 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1358 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
1359 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1360 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1361 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1362 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1363 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1364
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1365 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
1366 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1367 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1368 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1369 dst[j] = (dst[j] + ((683*(src[j] + 2*src[j+stride] + 1)) >> 11) + 1) >> 1;
1319
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1370 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1371 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1372 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1373 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1374 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1375
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1376 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
1377 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1378 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1379 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1380 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
1381 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1382 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1383 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1384 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1385 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1386
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1387 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
1388 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1389 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1390 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1391 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;
1319
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1392 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1393 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1394 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1395 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1396 }
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
1397 #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
1398 #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
1399 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
1400 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
1401 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
1402 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
1403 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
1404 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
1405 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
1406 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
1407 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
1408 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
1409 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
1410 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
1411 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
1412 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
1413 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
1414 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
1415 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
1416 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
1417 #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
1418
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1419 #define H264_CHROMA_MC(OPNAME, OP)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1420 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
1421 const int A=(8-x)*(8-y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1422 const int B=( x)*(8-y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1423 const int C=(8-x)*( y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1424 const int D=( x)*( y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1425 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1426 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1427 assert(x<8 && y<8 && x>=0 && y>=0);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1428 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1429 for(i=0; i<h; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1430 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1431 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
1432 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
1433 dst+= stride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1434 src+= stride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1435 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1436 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1437 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1438 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
1439 const int A=(8-x)*(8-y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1440 const int B=( x)*(8-y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1441 const int C=(8-x)*( y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1442 const int D=( x)*( y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1443 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1444 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1445 assert(x<8 && y<8 && x>=0 && y>=0);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1446 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1447 for(i=0; i<h; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1448 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1449 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
1450 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
1451 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
1452 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
1453 dst+= stride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1454 src+= stride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1455 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1456 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1457 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1458 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
1459 const int A=(8-x)*(8-y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1460 const int B=( x)*(8-y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1461 const int C=(8-x)*( y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1462 const int D=( x)*( y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1463 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1464 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1465 assert(x<8 && y<8 && x>=0 && y>=0);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1466 \
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 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
1470 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
1471 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
1472 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
1473 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
1474 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
1475 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
1476 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
1477 dst+= stride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1478 src+= stride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1479 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1480 }
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1481
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1482 #define op_avg(a, b) a = (((a)+(((b) + 32)>>6)+1)>>1)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1483 #define op_put(a, b) a = (((b) + 32)>>6)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1484
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1485 H264_CHROMA_MC(put_ , op_put)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1486 H264_CHROMA_MC(avg_ , op_avg)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1487 #undef op_avg
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1488 #undef op_put
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1489
3663
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1490 static void put_no_rnd_h264_chroma_mc8_c(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y){
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1491 const int A=(8-x)*(8-y);
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1492 const int B=( x)*(8-y);
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1493 const int C=(8-x)*( y);
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1494 const int D=( x)*( y);
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1495 int i;
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1496
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1497 assert(x<8 && y<8 && x>=0 && y>=0);
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1498
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1499 for(i=0; i<h; i++)
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1500 {
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1501 dst[0] = (A*src[0] + B*src[1] + C*src[stride+0] + D*src[stride+1] + 32 - 4) >> 6;
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1502 dst[1] = (A*src[1] + B*src[2] + C*src[stride+1] + D*src[stride+2] + 32 - 4) >> 6;
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1503 dst[2] = (A*src[2] + B*src[3] + C*src[stride+2] + D*src[stride+3] + 32 - 4) >> 6;
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1504 dst[3] = (A*src[3] + B*src[4] + C*src[stride+3] + D*src[stride+4] + 32 - 4) >> 6;
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1505 dst[4] = (A*src[4] + B*src[5] + C*src[stride+4] + D*src[stride+5] + 32 - 4) >> 6;
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1506 dst[5] = (A*src[5] + B*src[6] + C*src[stride+5] + D*src[stride+6] + 32 - 4) >> 6;
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1507 dst[6] = (A*src[6] + B*src[7] + C*src[stride+6] + D*src[stride+7] + 32 - 4) >> 6;
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1508 dst[7] = (A*src[7] + B*src[8] + C*src[stride+7] + D*src[stride+8] + 32 - 4) >> 6;
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1509 dst+= stride;
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1510 src+= stride;
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1511 }
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1512 }
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1513
3020
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
1514 static inline void copy_block2(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h)
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
1515 {
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
1516 int i;
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
1517 for(i=0; i<h; i++)
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
1518 {
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
1519 ST16(dst , LD16(src ));
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
1520 dst+=dstStride;
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
1521 src+=srcStride;
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
1522 }
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
1523 }
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
1524
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1525 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
1526 {
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1527 int i;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1528 for(i=0; i<h; i++)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1529 {
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1530 ST32(dst , LD32(src ));
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1531 dst+=dstStride;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1532 src+=srcStride;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1533 }
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1534 }
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1535
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1536 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
1537 {
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1538 int i;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1539 for(i=0; i<h; i++)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1540 {
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1541 ST32(dst , LD32(src ));
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1542 ST32(dst+4 , LD32(src+4 ));
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1543 dst+=dstStride;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1544 src+=srcStride;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1545 }
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1546 }
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1547
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1548 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
1549 {
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1550 int i;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1551 for(i=0; i<h; i++)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1552 {
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1553 ST32(dst , LD32(src ));
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1554 ST32(dst+4 , LD32(src+4 ));
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1555 ST32(dst+8 , LD32(src+8 ));
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1556 ST32(dst+12, LD32(src+12));
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1557 dst+=dstStride;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1558 src+=srcStride;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1559 }
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1560 }
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1561
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1562 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
1563 {
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1564 int i;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1565 for(i=0; i<h; i++)
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1566 {
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1567 ST32(dst , LD32(src ));
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1568 ST32(dst+4 , LD32(src+4 ));
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1569 ST32(dst+8 , LD32(src+8 ));
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1570 ST32(dst+12, LD32(src+12));
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1571 dst[16]= src[16];
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1572 dst+=dstStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1573 src+=srcStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1574 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1575 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1576
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1577 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
1578 {
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1579 int i;
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1580 for(i=0; i<h; i++)
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1581 {
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1582 ST32(dst , LD32(src ));
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1583 ST32(dst+4 , LD32(src+4 ));
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1584 dst[8]= src[8];
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1585 dst+=dstStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1586 src+=srcStride;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1587 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1588 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1589
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1590
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1591 #define QPEL_MC(r, OPNAME, RND, OP) \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1592 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
1593 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1594 int i;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1595 for(i=0; i<h; i++)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1596 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1597 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
1598 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
1599 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
1600 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
1601 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
1602 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
1603 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
1604 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
1605 dst+=dstStride;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1606 src+=srcStride;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1607 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1608 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1609 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1610 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
1611 const int w=8;\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1612 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1613 int i;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1614 for(i=0; i<w; i++)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1615 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1616 const int src0= src[0*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1617 const int src1= src[1*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1618 const int src2= src[2*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1619 const int src3= src[3*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1620 const int src4= src[4*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1621 const int src5= src[5*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1622 const int src6= src[6*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1623 const int src7= src[7*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1624 const int src8= src[8*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1625 OP(dst[0*dstStride], (src0+src1)*20 - (src0+src2)*6 + (src1+src3)*3 - (src2+src4));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1626 OP(dst[1*dstStride], (src1+src2)*20 - (src0+src3)*6 + (src0+src4)*3 - (src1+src5));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1627 OP(dst[2*dstStride], (src2+src3)*20 - (src1+src4)*6 + (src0+src5)*3 - (src0+src6));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1628 OP(dst[3*dstStride], (src3+src4)*20 - (src2+src5)*6 + (src1+src6)*3 - (src0+src7));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1629 OP(dst[4*dstStride], (src4+src5)*20 - (src3+src6)*6 + (src2+src7)*3 - (src1+src8));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1630 OP(dst[5*dstStride], (src5+src6)*20 - (src4+src7)*6 + (src3+src8)*3 - (src2+src8));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1631 OP(dst[6*dstStride], (src6+src7)*20 - (src5+src8)*6 + (src4+src8)*3 - (src3+src7));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1632 OP(dst[7*dstStride], (src7+src8)*20 - (src6+src8)*6 + (src5+src7)*3 - (src4+src6));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1633 dst++;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1634 src++;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1635 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1636 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1637 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1638 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
1639 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1640 int i;\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1641 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1642 for(i=0; i<h; i++)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1643 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1644 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
1645 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
1646 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
1647 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
1648 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
1649 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
1650 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
1651 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
1652 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
1653 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
1654 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
1655 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
1656 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
1657 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
1658 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
1659 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
1660 dst+=dstStride;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1661 src+=srcStride;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1662 }\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1663 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1664 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1665 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
1666 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1667 int i;\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1668 const int w=16;\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1669 for(i=0; i<w; i++)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1670 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1671 const int src0= src[0*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1672 const int src1= src[1*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1673 const int src2= src[2*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1674 const int src3= src[3*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1675 const int src4= src[4*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1676 const int src5= src[5*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1677 const int src6= src[6*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1678 const int src7= src[7*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1679 const int src8= src[8*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1680 const int src9= src[9*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1681 const int src10= src[10*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1682 const int src11= src[11*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1683 const int src12= src[12*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1684 const int src13= src[13*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1685 const int src14= src[14*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1686 const int src15= src[15*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1687 const int src16= src[16*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1688 OP(dst[ 0*dstStride], (src0 +src1 )*20 - (src0 +src2 )*6 + (src1 +src3 )*3 - (src2 +src4 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1689 OP(dst[ 1*dstStride], (src1 +src2 )*20 - (src0 +src3 )*6 + (src0 +src4 )*3 - (src1 +src5 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1690 OP(dst[ 2*dstStride], (src2 +src3 )*20 - (src1 +src4 )*6 + (src0 +src5 )*3 - (src0 +src6 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1691 OP(dst[ 3*dstStride], (src3 +src4 )*20 - (src2 +src5 )*6 + (src1 +src6 )*3 - (src0 +src7 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1692 OP(dst[ 4*dstStride], (src4 +src5 )*20 - (src3 +src6 )*6 + (src2 +src7 )*3 - (src1 +src8 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1693 OP(dst[ 5*dstStride], (src5 +src6 )*20 - (src4 +src7 )*6 + (src3 +src8 )*3 - (src2 +src9 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1694 OP(dst[ 6*dstStride], (src6 +src7 )*20 - (src5 +src8 )*6 + (src4 +src9 )*3 - (src3 +src10));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1695 OP(dst[ 7*dstStride], (src7 +src8 )*20 - (src6 +src9 )*6 + (src5 +src10)*3 - (src4 +src11));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1696 OP(dst[ 8*dstStride], (src8 +src9 )*20 - (src7 +src10)*6 + (src6 +src11)*3 - (src5 +src12));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1697 OP(dst[ 9*dstStride], (src9 +src10)*20 - (src8 +src11)*6 + (src7 +src12)*3 - (src6 +src13));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1698 OP(dst[10*dstStride], (src10+src11)*20 - (src9 +src12)*6 + (src8 +src13)*3 - (src7 +src14));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1699 OP(dst[11*dstStride], (src11+src12)*20 - (src10+src13)*6 + (src9 +src14)*3 - (src8 +src15));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1700 OP(dst[12*dstStride], (src12+src13)*20 - (src11+src14)*6 + (src10+src15)*3 - (src9 +src16));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1701 OP(dst[13*dstStride], (src13+src14)*20 - (src12+src15)*6 + (src11+src16)*3 - (src10+src16));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1702 OP(dst[14*dstStride], (src14+src15)*20 - (src13+src16)*6 + (src12+src16)*3 - (src11+src15));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1703 OP(dst[15*dstStride], (src15+src16)*20 - (src14+src16)*6 + (src13+src15)*3 - (src12+src14));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1704 dst++;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1705 src++;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1706 }\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1707 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1708 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1709 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
1710 OPNAME ## pixels8_c(dst, src, stride, 8);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1711 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1712 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1713 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
1714 uint8_t half[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1715 put ## RND ## mpeg4_qpel8_h_lowpass(half, src, 8, stride, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1716 OPNAME ## pixels8_l2(dst, src, half, stride, stride, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1717 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1718 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1719 static void OPNAME ## qpel8_mc20_c(uint8_t *dst, uint8_t *src, int stride){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1720 OPNAME ## mpeg4_qpel8_h_lowpass(dst, src, stride, stride, 8);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1721 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1722 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1723 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
1724 uint8_t half[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1725 put ## RND ## mpeg4_qpel8_h_lowpass(half, src, 8, stride, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1726 OPNAME ## pixels8_l2(dst, src+1, half, stride, stride, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1727 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1728 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1729 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
1730 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1731 uint8_t half[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1732 copy_block9(full, src, 16, stride, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1733 put ## RND ## mpeg4_qpel8_v_lowpass(half, full, 8, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1734 OPNAME ## pixels8_l2(dst, full, half, stride, 16, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1735 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1736 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1737 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
1738 uint8_t full[16*9];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1739 copy_block9(full, src, 16, stride, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1740 OPNAME ## mpeg4_qpel8_v_lowpass(dst, full, stride, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1741 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1742 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1743 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
1744 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1745 uint8_t half[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1746 copy_block9(full, src, 16, stride, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1747 put ## RND ## mpeg4_qpel8_v_lowpass(half, full, 8, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1748 OPNAME ## pixels8_l2(dst, full+16, half, stride, 16, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1749 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1750 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
1751 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1752 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1753 uint8_t halfV[64];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1754 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1755 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1756 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1757 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full, 8, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1758 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1759 OPNAME ## pixels8_l4(dst, full, halfH, halfV, halfHV, stride, 16, 8, 8, 8, 8);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1760 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1761 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
1762 uint8_t full[16*9];\
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];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1765 copy_block9(full, src, 16, stride, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1766 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1767 put ## RND ## pixels8_l2(halfH, halfH, full, 8, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1768 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1769 OPNAME ## pixels8_l2(dst, halfH, halfHV, stride, 8, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1770 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1771 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
1772 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1773 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1774 uint8_t halfV[64];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1775 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1776 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1777 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1778 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full+1, 8, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1779 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1780 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
1781 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1782 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
1783 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1784 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1785 uint8_t halfHV[64];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1786 copy_block9(full, src, 16, stride, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1787 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1788 put ## RND ## pixels8_l2(halfH, halfH, full+1, 8, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1789 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1790 OPNAME ## pixels8_l2(dst, halfH, halfHV, stride, 8, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1791 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1792 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
1793 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1794 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1795 uint8_t halfV[64];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1796 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1797 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1798 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1799 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full, 8, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1800 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1801 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
1802 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1803 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
1804 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1805 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1806 uint8_t halfHV[64];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1807 copy_block9(full, src, 16, stride, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1808 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1809 put ## RND ## pixels8_l2(halfH, halfH, full, 8, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1810 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1811 OPNAME ## pixels8_l2(dst, halfH+8, halfHV, stride, 8, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1812 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1813 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
1814 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1815 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1816 uint8_t halfV[64];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1817 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1818 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1819 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full , 8, 16, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1820 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full+1, 8, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1821 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1822 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
1823 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1824 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
1825 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1826 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1827 uint8_t halfHV[64];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1828 copy_block9(full, src, 16, stride, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1829 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1830 put ## RND ## pixels8_l2(halfH, halfH, full+1, 8, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1831 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1832 OPNAME ## pixels8_l2(dst, halfH+8, halfHV, stride, 8, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1833 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1834 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
1835 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1836 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1837 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, src, 8, stride, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1838 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1839 OPNAME ## pixels8_l2(dst, halfH, halfHV, stride, 8, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1840 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1841 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
1842 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1843 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1844 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, src, 8, stride, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1845 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1846 OPNAME ## pixels8_l2(dst, halfH+8, halfHV, stride, 8, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1847 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1848 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
1849 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1850 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1851 uint8_t halfV[64];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1852 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1853 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1854 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1855 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full, 8, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1856 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1857 OPNAME ## pixels8_l2(dst, halfV, halfHV, stride, 8, 8, 8);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1858 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1859 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
1860 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1861 uint8_t halfH[72];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1862 copy_block9(full, src, 16, stride, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1863 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1864 put ## RND ## pixels8_l2(halfH, halfH, full, 8, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1865 OPNAME ## mpeg4_qpel8_v_lowpass(dst, halfH, stride, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1866 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1867 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
1868 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1869 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1870 uint8_t halfV[64];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1871 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1872 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1873 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1874 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full+1, 8, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1875 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1876 OPNAME ## pixels8_l2(dst, halfV, halfHV, stride, 8, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1877 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1878 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
1879 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1880 uint8_t halfH[72];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1881 copy_block9(full, src, 16, stride, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1882 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1883 put ## RND ## pixels8_l2(halfH, halfH, full+1, 8, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1884 OPNAME ## mpeg4_qpel8_v_lowpass(dst, halfH, stride, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1885 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1886 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
1887 uint8_t halfH[72];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1888 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, src, 8, stride, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1889 OPNAME ## mpeg4_qpel8_v_lowpass(dst, halfH, stride, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1890 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1891 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
1892 OPNAME ## pixels16_c(dst, src, stride, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1893 }\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1894 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1895 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
1896 uint8_t half[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1897 put ## RND ## mpeg4_qpel16_h_lowpass(half, src, 16, stride, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1898 OPNAME ## pixels16_l2(dst, src, half, stride, stride, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1899 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1900 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1901 static void OPNAME ## qpel16_mc20_c(uint8_t *dst, uint8_t *src, int stride){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1902 OPNAME ## mpeg4_qpel16_h_lowpass(dst, src, stride, stride, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1903 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1904 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1905 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
1906 uint8_t half[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1907 put ## RND ## mpeg4_qpel16_h_lowpass(half, src, 16, stride, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1908 OPNAME ## pixels16_l2(dst, src+1, half, stride, stride, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1909 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1910 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1911 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
1912 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1913 uint8_t half[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1914 copy_block17(full, src, 24, stride, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1915 put ## RND ## mpeg4_qpel16_v_lowpass(half, full, 16, 24);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1916 OPNAME ## pixels16_l2(dst, full, half, stride, 24, 16, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1917 }\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1918 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1919 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
1920 uint8_t full[24*17];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1921 copy_block17(full, src, 24, stride, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1922 OPNAME ## mpeg4_qpel16_v_lowpass(dst, full, stride, 24);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1923 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1924 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1925 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
1926 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1927 uint8_t half[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1928 copy_block17(full, src, 24, stride, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1929 put ## RND ## mpeg4_qpel16_v_lowpass(half, full, 16, 24);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1930 OPNAME ## pixels16_l2(dst, full+24, half, stride, 24, 16, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1931 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1932 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
1933 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1934 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1935 uint8_t halfV[256];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1936 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1937 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1938 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1939 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full, 16, 24);\
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1940 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1941 OPNAME ## pixels16_l4(dst, full, halfH, halfV, halfHV, stride, 24, 16, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1942 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1943 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
1944 uint8_t full[24*17];\
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];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1947 copy_block17(full, src, 24, stride, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1948 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1949 put ## RND ## pixels16_l2(halfH, halfH, full, 16, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1950 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1951 OPNAME ## pixels16_l2(dst, halfH, halfHV, stride, 16, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1952 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1953 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
1954 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1955 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1956 uint8_t halfV[256];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1957 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1958 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1959 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1960 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full+1, 16, 24);\
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1961 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1962 OPNAME ## pixels16_l4(dst, full+1, halfH, halfV, halfHV, stride, 24, 16, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1963 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1964 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
1965 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1966 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1967 uint8_t halfHV[256];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1968 copy_block17(full, src, 24, stride, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1969 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1970 put ## RND ## pixels16_l2(halfH, halfH, full+1, 16, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1971 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1972 OPNAME ## pixels16_l2(dst, halfH, halfHV, stride, 16, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1973 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1974 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
1975 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1976 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1977 uint8_t halfV[256];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1978 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1979 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1980 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1981 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full, 16, 24);\
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1982 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1983 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
1984 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1985 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
1986 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1987 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1988 uint8_t halfHV[256];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1989 copy_block17(full, src, 24, stride, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1990 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1991 put ## RND ## pixels16_l2(halfH, halfH, full, 16, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1992 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1993 OPNAME ## pixels16_l2(dst, halfH+16, halfHV, stride, 16, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1994 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1995 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
1996 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1997 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1998 uint8_t halfV[256];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1999 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2000 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2001 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full , 16, 24, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2002 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full+1, 16, 24);\
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2003 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2004 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
2005 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2006 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
2007 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2008 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2009 uint8_t halfHV[256];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2010 copy_block17(full, src, 24, stride, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2011 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2012 put ## RND ## pixels16_l2(halfH, halfH, full+1, 16, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2013 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2014 OPNAME ## pixels16_l2(dst, halfH+16, halfHV, stride, 16, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2015 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2016 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
2017 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2018 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2019 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, src, 16, stride, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2020 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2021 OPNAME ## pixels16_l2(dst, halfH, halfHV, stride, 16, 16, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
2022 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2023 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
2024 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2025 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2026 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, src, 16, stride, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2027 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2028 OPNAME ## pixels16_l2(dst, halfH+16, halfHV, stride, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2029 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2030 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
2031 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2032 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2033 uint8_t halfV[256];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2034 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2035 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2036 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2037 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full, 16, 24);\
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2038 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2039 OPNAME ## pixels16_l2(dst, halfV, halfHV, stride, 16, 16, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
2040 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2041 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
2042 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2043 uint8_t halfH[272];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2044 copy_block17(full, src, 24, stride, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2045 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2046 put ## RND ## pixels16_l2(halfH, halfH, full, 16, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2047 OPNAME ## mpeg4_qpel16_v_lowpass(dst, halfH, stride, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2048 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2049 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
2050 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2051 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2052 uint8_t halfV[256];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2053 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2054 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2055 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2056 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full+1, 16, 24);\
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2057 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2058 OPNAME ## pixels16_l2(dst, halfV, halfHV, stride, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2059 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2060 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
2061 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2062 uint8_t halfH[272];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2063 copy_block17(full, src, 24, stride, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2064 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2065 put ## RND ## pixels16_l2(halfH, halfH, full+1, 16, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2066 OPNAME ## mpeg4_qpel16_v_lowpass(dst, halfH, stride, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2067 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2068 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
2069 uint8_t halfH[272];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2070 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, src, 16, stride, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2071 OPNAME ## mpeg4_qpel16_v_lowpass(dst, halfH, stride, 16);\
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
2072 }
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
2073
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2074 #define op_avg(a, b) a = (((a)+cm[((b) + 16)>>5]+1)>>1)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2075 #define op_avg_no_rnd(a, b) a = (((a)+cm[((b) + 15)>>5])>>1)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2076 #define op_put(a, b) a = cm[((b) + 16)>>5]
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2077 #define op_put_no_rnd(a, b) a = cm[((b) + 15)>>5]
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2078
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2079 QPEL_MC(0, put_ , _ , op_put)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2080 QPEL_MC(1, put_no_rnd_, _no_rnd_, op_put_no_rnd)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2081 QPEL_MC(0, avg_ , _ , op_avg)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2082 //QPEL_MC(1, avg_no_rnd , _ , op_avg)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2083 #undef op_avg
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2084 #undef op_avg_no_rnd
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2085 #undef op_put
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2086 #undef op_put_no_rnd
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
2087
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2088 #if 1
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2089 #define H264_LOWPASS(OPNAME, OP, OP2) \
3020
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2090 static void OPNAME ## h264_qpel2_h_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2091 const int h=2;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2092 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2093 int i;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2094 for(i=0; i<h; i++)\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2095 {\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2096 OP(dst[0], (src[0]+src[1])*20 - (src[-1]+src[2])*5 + (src[-2]+src[3]));\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2097 OP(dst[1], (src[1]+src[2])*20 - (src[0 ]+src[3])*5 + (src[-1]+src[4]));\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2098 dst+=dstStride;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2099 src+=srcStride;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2100 }\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2101 }\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2102 \
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2103 static void OPNAME ## h264_qpel2_v_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2104 const int w=2;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2105 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2106 int i;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2107 for(i=0; i<w; i++)\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2108 {\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2109 const int srcB= src[-2*srcStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2110 const int srcA= src[-1*srcStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2111 const int src0= src[0 *srcStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2112 const int src1= src[1 *srcStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2113 const int src2= src[2 *srcStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2114 const int src3= src[3 *srcStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2115 const int src4= src[4 *srcStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2116 OP(dst[0*dstStride], (src0+src1)*20 - (srcA+src2)*5 + (srcB+src3));\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2117 OP(dst[1*dstStride], (src1+src2)*20 - (src0+src3)*5 + (srcA+src4));\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2118 dst++;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2119 src++;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2120 }\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2121 }\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2122 \
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2123 static void OPNAME ## h264_qpel2_hv_lowpass(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2124 const int h=2;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2125 const int w=2;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2126 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2127 int i;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2128 src -= 2*srcStride;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2129 for(i=0; i<h+5; i++)\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2130 {\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2131 tmp[0]= (src[0]+src[1])*20 - (src[-1]+src[2])*5 + (src[-2]+src[3]);\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2132 tmp[1]= (src[1]+src[2])*20 - (src[0 ]+src[3])*5 + (src[-1]+src[4]);\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2133 tmp+=tmpStride;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2134 src+=srcStride;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2135 }\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2136 tmp -= tmpStride*(h+5-2);\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2137 for(i=0; i<w; i++)\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2138 {\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2139 const int tmpB= tmp[-2*tmpStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2140 const int tmpA= tmp[-1*tmpStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2141 const int tmp0= tmp[0 *tmpStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2142 const int tmp1= tmp[1 *tmpStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2143 const int tmp2= tmp[2 *tmpStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2144 const int tmp3= tmp[3 *tmpStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2145 const int tmp4= tmp[4 *tmpStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2146 OP2(dst[0*dstStride], (tmp0+tmp1)*20 - (tmpA+tmp2)*5 + (tmpB+tmp3));\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2147 OP2(dst[1*dstStride], (tmp1+tmp2)*20 - (tmp0+tmp3)*5 + (tmpA+tmp4));\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2148 dst++;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2149 tmp++;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2150 }\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2151 }\
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2152 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
2153 const int h=4;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2154 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2155 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2156 for(i=0; i<h; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2157 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2158 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
2159 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
2160 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
2161 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
2162 dst+=dstStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2163 src+=srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2164 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2165 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2166 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2167 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
2168 const int w=4;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2169 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2170 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2171 for(i=0; i<w; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2172 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2173 const int srcB= src[-2*srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2174 const int srcA= src[-1*srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2175 const int src0= src[0 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2176 const int src1= src[1 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2177 const int src2= src[2 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2178 const int src3= src[3 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2179 const int src4= src[4 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2180 const int src5= src[5 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2181 const int src6= src[6 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2182 OP(dst[0*dstStride], (src0+src1)*20 - (srcA+src2)*5 + (srcB+src3));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2183 OP(dst[1*dstStride], (src1+src2)*20 - (src0+src3)*5 + (srcA+src4));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2184 OP(dst[2*dstStride], (src2+src3)*20 - (src1+src4)*5 + (src0+src5));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2185 OP(dst[3*dstStride], (src3+src4)*20 - (src2+src5)*5 + (src1+src6));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2186 dst++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2187 src++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2188 }\
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 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
2192 const int h=4;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2193 const int w=4;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2194 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2195 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2196 src -= 2*srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2197 for(i=0; i<h+5; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2198 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2199 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
2200 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
2201 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
2202 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
2203 tmp+=tmpStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2204 src+=srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2205 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2206 tmp -= tmpStride*(h+5-2);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2207 for(i=0; i<w; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2208 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2209 const int tmpB= tmp[-2*tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2210 const int tmpA= tmp[-1*tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2211 const int tmp0= tmp[0 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2212 const int tmp1= tmp[1 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2213 const int tmp2= tmp[2 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2214 const int tmp3= tmp[3 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2215 const int tmp4= tmp[4 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2216 const int tmp5= tmp[5 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2217 const int tmp6= tmp[6 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2218 OP2(dst[0*dstStride], (tmp0+tmp1)*20 - (tmpA+tmp2)*5 + (tmpB+tmp3));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2219 OP2(dst[1*dstStride], (tmp1+tmp2)*20 - (tmp0+tmp3)*5 + (tmpA+tmp4));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2220 OP2(dst[2*dstStride], (tmp2+tmp3)*20 - (tmp1+tmp4)*5 + (tmp0+tmp5));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2221 OP2(dst[3*dstStride], (tmp3+tmp4)*20 - (tmp2+tmp5)*5 + (tmp1+tmp6));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2222 dst++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2223 tmp++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2224 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2225 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2226 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2227 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
2228 const int h=8;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2229 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2230 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2231 for(i=0; i<h; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2232 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2233 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
2234 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
2235 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
2236 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
2237 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
2238 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
2239 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
2240 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
2241 dst+=dstStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2242 src+=srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2243 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2244 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2245 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2246 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
2247 const int w=8;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2248 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2249 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2250 for(i=0; i<w; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2251 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2252 const int srcB= src[-2*srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2253 const int srcA= src[-1*srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2254 const int src0= src[0 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2255 const int src1= src[1 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2256 const int src2= src[2 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2257 const int src3= src[3 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2258 const int src4= src[4 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2259 const int src5= src[5 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2260 const int src6= src[6 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2261 const int src7= src[7 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2262 const int src8= src[8 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2263 const int src9= src[9 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2264 const int src10=src[10*srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2265 OP(dst[0*dstStride], (src0+src1)*20 - (srcA+src2)*5 + (srcB+src3));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2266 OP(dst[1*dstStride], (src1+src2)*20 - (src0+src3)*5 + (srcA+src4));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2267 OP(dst[2*dstStride], (src2+src3)*20 - (src1+src4)*5 + (src0+src5));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2268 OP(dst[3*dstStride], (src3+src4)*20 - (src2+src5)*5 + (src1+src6));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2269 OP(dst[4*dstStride], (src4+src5)*20 - (src3+src6)*5 + (src2+src7));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2270 OP(dst[5*dstStride], (src5+src6)*20 - (src4+src7)*5 + (src3+src8));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2271 OP(dst[6*dstStride], (src6+src7)*20 - (src5+src8)*5 + (src4+src9));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2272 OP(dst[7*dstStride], (src7+src8)*20 - (src6+src9)*5 + (src5+src10));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2273 dst++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2274 src++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2275 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2276 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2277 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2278 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
2279 const int h=8;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2280 const int w=8;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2281 uint8_t *cm = cropTbl + MAX_NEG_CROP;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2282 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2283 src -= 2*srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2284 for(i=0; i<h+5; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2285 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2286 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
2287 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
2288 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
2289 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
2290 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
2291 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
2292 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
2293 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
2294 tmp+=tmpStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2295 src+=srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2296 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2297 tmp -= tmpStride*(h+5-2);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2298 for(i=0; i<w; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2299 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2300 const int tmpB= tmp[-2*tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2301 const int tmpA= tmp[-1*tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2302 const int tmp0= tmp[0 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2303 const int tmp1= tmp[1 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2304 const int tmp2= tmp[2 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2305 const int tmp3= tmp[3 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2306 const int tmp4= tmp[4 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2307 const int tmp5= tmp[5 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2308 const int tmp6= tmp[6 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2309 const int tmp7= tmp[7 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2310 const int tmp8= tmp[8 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2311 const int tmp9= tmp[9 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2312 const int tmp10=tmp[10*tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2313 OP2(dst[0*dstStride], (tmp0+tmp1)*20 - (tmpA+tmp2)*5 + (tmpB+tmp3));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2314 OP2(dst[1*dstStride], (tmp1+tmp2)*20 - (tmp0+tmp3)*5 + (tmpA+tmp4));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2315 OP2(dst[2*dstStride], (tmp2+tmp3)*20 - (tmp1+tmp4)*5 + (tmp0+tmp5));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2316 OP2(dst[3*dstStride], (tmp3+tmp4)*20 - (tmp2+tmp5)*5 + (tmp1+tmp6));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2317 OP2(dst[4*dstStride], (tmp4+tmp5)*20 - (tmp3+tmp6)*5 + (tmp2+tmp7));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2318 OP2(dst[5*dstStride], (tmp5+tmp6)*20 - (tmp4+tmp7)*5 + (tmp3+tmp8));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2319 OP2(dst[6*dstStride], (tmp6+tmp7)*20 - (tmp5+tmp8)*5 + (tmp4+tmp9));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2320 OP2(dst[7*dstStride], (tmp7+tmp8)*20 - (tmp6+tmp9)*5 + (tmp5+tmp10));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2321 dst++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2322 tmp++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2323 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2324 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2325 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2326 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
2327 OPNAME ## h264_qpel8_v_lowpass(dst , src , dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2328 OPNAME ## h264_qpel8_v_lowpass(dst+8, src+8, dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2329 src += 8*srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2330 dst += 8*dstStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2331 OPNAME ## h264_qpel8_v_lowpass(dst , src , dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2332 OPNAME ## h264_qpel8_v_lowpass(dst+8, src+8, dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2333 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2334 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2335 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
2336 OPNAME ## h264_qpel8_h_lowpass(dst , src , dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2337 OPNAME ## h264_qpel8_h_lowpass(dst+8, src+8, dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2338 src += 8*srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2339 dst += 8*dstStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2340 OPNAME ## h264_qpel8_h_lowpass(dst , src , dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2341 OPNAME ## h264_qpel8_h_lowpass(dst+8, src+8, dstStride, srcStride);\
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_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
2345 OPNAME ## h264_qpel8_hv_lowpass(dst , tmp , src , dstStride, tmpStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2346 OPNAME ## h264_qpel8_hv_lowpass(dst+8, tmp+8, src+8, dstStride, tmpStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2347 src += 8*srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2348 dst += 8*dstStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2349 OPNAME ## h264_qpel8_hv_lowpass(dst , tmp , src , dstStride, tmpStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2350 OPNAME ## h264_qpel8_hv_lowpass(dst+8, tmp+8, src+8, dstStride, tmpStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2351 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2352
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2353 #define H264_MC(OPNAME, SIZE) \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2354 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
2355 OPNAME ## pixels ## SIZE ## _c(dst, src, stride, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2356 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2357 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2358 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
2359 uint8_t half[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2360 put_h264_qpel ## SIZE ## _h_lowpass(half, src, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2361 OPNAME ## pixels ## SIZE ## _l2(dst, src, half, stride, stride, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2362 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2363 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2364 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
2365 OPNAME ## h264_qpel ## SIZE ## _h_lowpass(dst, src, stride, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2366 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2367 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2368 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
2369 uint8_t half[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2370 put_h264_qpel ## SIZE ## _h_lowpass(half, src, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2371 OPNAME ## pixels ## SIZE ## _l2(dst, src+1, half, stride, stride, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2372 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2373 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2374 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
2375 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2376 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2377 uint8_t half[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2378 copy_block ## SIZE (full, src - stride*2, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2379 put_h264_qpel ## SIZE ## _v_lowpass(half, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2380 OPNAME ## pixels ## SIZE ## _l2(dst, full_mid, half, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2381 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2382 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2383 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
2384 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2385 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2386 copy_block ## SIZE (full, src - stride*2, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2387 OPNAME ## h264_qpel ## SIZE ## _v_lowpass(dst, full_mid, stride, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2388 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2389 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2390 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
2391 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2392 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2393 uint8_t half[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2394 copy_block ## SIZE (full, src - stride*2, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2395 put_h264_qpel ## SIZE ## _v_lowpass(half, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2396 OPNAME ## pixels ## SIZE ## _l2(dst, full_mid+SIZE, half, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2397 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2398 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2399 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
2400 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2401 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2402 uint8_t halfH[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2403 uint8_t halfV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2404 put_h264_qpel ## SIZE ## _h_lowpass(halfH, src, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2405 copy_block ## SIZE (full, src - stride*2, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2406 put_h264_qpel ## SIZE ## _v_lowpass(halfV, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2407 OPNAME ## pixels ## SIZE ## _l2(dst, halfH, halfV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2408 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2409 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2410 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
2411 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2412 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2413 uint8_t halfH[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2414 uint8_t halfV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2415 put_h264_qpel ## SIZE ## _h_lowpass(halfH, src, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2416 copy_block ## SIZE (full, src - stride*2 + 1, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2417 put_h264_qpel ## SIZE ## _v_lowpass(halfV, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2418 OPNAME ## pixels ## SIZE ## _l2(dst, halfH, halfV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2419 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2420 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2421 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
2422 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2423 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2424 uint8_t halfH[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2425 uint8_t halfV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2426 put_h264_qpel ## SIZE ## _h_lowpass(halfH, src + stride, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2427 copy_block ## SIZE (full, src - stride*2, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2428 put_h264_qpel ## SIZE ## _v_lowpass(halfV, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2429 OPNAME ## pixels ## SIZE ## _l2(dst, halfH, halfV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2430 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2431 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2432 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
2433 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2434 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2435 uint8_t halfH[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2436 uint8_t halfV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2437 put_h264_qpel ## SIZE ## _h_lowpass(halfH, src + stride, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2438 copy_block ## SIZE (full, src - stride*2 + 1, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2439 put_h264_qpel ## SIZE ## _v_lowpass(halfV, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2440 OPNAME ## pixels ## SIZE ## _l2(dst, halfH, halfV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2441 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2442 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2443 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
2444 int16_t tmp[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2445 OPNAME ## h264_qpel ## SIZE ## _hv_lowpass(dst, tmp, src, stride, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2446 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2447 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2448 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
2449 int16_t tmp[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2450 uint8_t halfH[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2451 uint8_t halfHV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2452 put_h264_qpel ## SIZE ## _h_lowpass(halfH, src, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2453 put_h264_qpel ## SIZE ## _hv_lowpass(halfHV, tmp, src, SIZE, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2454 OPNAME ## pixels ## SIZE ## _l2(dst, halfH, halfHV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2455 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2456 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2457 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
2458 int16_t tmp[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2459 uint8_t halfH[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2460 uint8_t halfHV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2461 put_h264_qpel ## SIZE ## _h_lowpass(halfH, src + stride, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2462 put_h264_qpel ## SIZE ## _hv_lowpass(halfHV, tmp, src, SIZE, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2463 OPNAME ## pixels ## SIZE ## _l2(dst, halfH, halfHV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2464 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2465 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2466 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
2467 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2468 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2469 int16_t tmp[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2470 uint8_t halfV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2471 uint8_t halfHV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2472 copy_block ## SIZE (full, src - stride*2, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2473 put_h264_qpel ## SIZE ## _v_lowpass(halfV, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2474 put_h264_qpel ## SIZE ## _hv_lowpass(halfHV, tmp, src, SIZE, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2475 OPNAME ## pixels ## SIZE ## _l2(dst, halfV, halfHV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2476 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2477 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2478 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
2479 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2480 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2481 int16_t tmp[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2482 uint8_t halfV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2483 uint8_t halfHV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2484 copy_block ## SIZE (full, src - stride*2 + 1, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2485 put_h264_qpel ## SIZE ## _v_lowpass(halfV, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2486 put_h264_qpel ## SIZE ## _hv_lowpass(halfHV, tmp, src, SIZE, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2487 OPNAME ## pixels ## SIZE ## _l2(dst, halfV, halfHV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2488 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2489
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2490 #define op_avg(a, b) a = (((a)+cm[((b) + 16)>>5]+1)>>1)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2491 //#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
2492 #define op_put(a, b) a = cm[((b) + 16)>>5]
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2493 #define op2_avg(a, b) a = (((a)+cm[((b) + 512)>>10]+1)>>1)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2494 #define op2_put(a, b) a = cm[((b) + 512)>>10]
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2495
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2496 H264_LOWPASS(put_ , op_put, op2_put)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2497 H264_LOWPASS(avg_ , op_avg, op2_avg)
3020
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2498 H264_MC(put_, 2)
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2499 H264_MC(put_, 4)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2500 H264_MC(put_, 8)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2501 H264_MC(put_, 16)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2502 H264_MC(avg_, 4)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2503 H264_MC(avg_, 8)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2504 H264_MC(avg_, 16)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2505
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2506 #undef op_avg
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2507 #undef op_put
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2508 #undef op2_avg
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2509 #undef op2_put
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2510 #endif
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2511
2448
bfa9192a22ce use clip_uint8()
michael
parents: 2415
diff changeset
2512 #define op_scale1(x) block[x] = clip_uint8( (block[x]*weight + offset) >> log2_denom )
bfa9192a22ce use clip_uint8()
michael
parents: 2415
diff changeset
2513 #define op_scale2(x) dst[x] = clip_uint8( (src[x]*weights + dst[x]*weightd + offset) >> (log2_denom+1))
2415
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2514 #define H264_WEIGHT(W,H) \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2515 static void weight_h264_pixels ## W ## x ## H ## _c(uint8_t *block, int stride, int log2_denom, int weight, int offset){ \
3029
7a92269a3ccd tweak h264_biweight
lorenm
parents: 3020
diff changeset
2516 int y; \
2415
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2517 offset <<= log2_denom; \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2518 if(log2_denom) offset += 1<<(log2_denom-1); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2519 for(y=0; y<H; y++, block += stride){ \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2520 op_scale1(0); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2521 op_scale1(1); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2522 if(W==2) continue; \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2523 op_scale1(2); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2524 op_scale1(3); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2525 if(W==4) continue; \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2526 op_scale1(4); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2527 op_scale1(5); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2528 op_scale1(6); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2529 op_scale1(7); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2530 if(W==8) continue; \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2531 op_scale1(8); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2532 op_scale1(9); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2533 op_scale1(10); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2534 op_scale1(11); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2535 op_scale1(12); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2536 op_scale1(13); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2537 op_scale1(14); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2538 op_scale1(15); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2539 } \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2540 } \
3029
7a92269a3ccd tweak h264_biweight
lorenm
parents: 3020
diff changeset
2541 static void biweight_h264_pixels ## W ## x ## H ## _c(uint8_t *dst, uint8_t *src, int stride, int log2_denom, int weightd, int weights, int offset){ \
7a92269a3ccd tweak h264_biweight
lorenm
parents: 3020
diff changeset
2542 int y; \
7a92269a3ccd tweak h264_biweight
lorenm
parents: 3020
diff changeset
2543 offset = ((offset + 1) | 1) << log2_denom; \
2415
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2544 for(y=0; y<H; y++, dst += stride, src += stride){ \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2545 op_scale2(0); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2546 op_scale2(1); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2547 if(W==2) continue; \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2548 op_scale2(2); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2549 op_scale2(3); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2550 if(W==4) continue; \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2551 op_scale2(4); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2552 op_scale2(5); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2553 op_scale2(6); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2554 op_scale2(7); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2555 if(W==8) continue; \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2556 op_scale2(8); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2557 op_scale2(9); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2558 op_scale2(10); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2559 op_scale2(11); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2560 op_scale2(12); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2561 op_scale2(13); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2562 op_scale2(14); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2563 op_scale2(15); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2564 } \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2565 }
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2566
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2567 H264_WEIGHT(16,16)
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2568 H264_WEIGHT(16,8)
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2569 H264_WEIGHT(8,16)
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2570 H264_WEIGHT(8,8)
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2571 H264_WEIGHT(8,4)
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2572 H264_WEIGHT(4,8)
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2573 H264_WEIGHT(4,4)
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2574 H264_WEIGHT(4,2)
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2575 H264_WEIGHT(2,4)
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2576 H264_WEIGHT(2,2)
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2577
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2578 #undef op_scale1
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2579 #undef op_scale2
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2580 #undef H264_WEIGHT
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2581
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2582 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
2583 uint8_t *cm = cropTbl + MAX_NEG_CROP;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2584 int i;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2585
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2586 for(i=0; i<h; i++){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2587 dst[0]= cm[(9*(src[0] + src[1]) - (src[-1] + src[2]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2588 dst[1]= cm[(9*(src[1] + src[2]) - (src[ 0] + src[3]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2589 dst[2]= cm[(9*(src[2] + src[3]) - (src[ 1] + src[4]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2590 dst[3]= cm[(9*(src[3] + src[4]) - (src[ 2] + src[5]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2591 dst[4]= cm[(9*(src[4] + src[5]) - (src[ 3] + src[6]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2592 dst[5]= cm[(9*(src[5] + src[6]) - (src[ 4] + src[7]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2593 dst[6]= cm[(9*(src[6] + src[7]) - (src[ 5] + src[8]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2594 dst[7]= cm[(9*(src[7] + src[8]) - (src[ 6] + src[9]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2595 dst+=dstStride;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2596 src+=srcStride;
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2597 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2598 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2599
3432
ab49baf4adad Fix compilation with --disable-decoders.
diego
parents: 3395
diff changeset
2600 #ifdef CONFIG_CAVS_DECODER
3395
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2601 /* AVS specific */
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2602 void ff_cavsdsp_init(DSPContext* c, AVCodecContext *avctx);
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2603
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2604 void ff_put_cavs_qpel8_mc00_c(uint8_t *dst, uint8_t *src, int stride) {
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2605 put_pixels8_c(dst, src, stride, 8);
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2606 }
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2607 void ff_avg_cavs_qpel8_mc00_c(uint8_t *dst, uint8_t *src, int stride) {
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2608 avg_pixels8_c(dst, src, stride, 8);
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2609 }
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2610 void ff_put_cavs_qpel16_mc00_c(uint8_t *dst, uint8_t *src, int stride) {
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2611 put_pixels16_c(dst, src, stride, 16);
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2612 }
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2613 void ff_avg_cavs_qpel16_mc00_c(uint8_t *dst, uint8_t *src, int stride) {
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2614 avg_pixels16_c(dst, src, stride, 16);
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2615 }
3432
ab49baf4adad Fix compilation with --disable-decoders.
diego
parents: 3395
diff changeset
2616 #endif /* CONFIG_CAVS_DECODER */
3395
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2617
3526
7dc8e4a12105 New functions in DSPContext for VC-1 decoding
kostya
parents: 3432
diff changeset
2618 #if defined(CONFIG_VC1_DECODER) || defined(CONFIG_WMV3_DECODER)
7dc8e4a12105 New functions in DSPContext for VC-1 decoding
kostya
parents: 3432
diff changeset
2619 /* VC-1 specific */
7dc8e4a12105 New functions in DSPContext for VC-1 decoding
kostya
parents: 3432
diff changeset
2620 void ff_vc1dsp_init(DSPContext* c, AVCodecContext *avctx);
7dc8e4a12105 New functions in DSPContext for VC-1 decoding
kostya
parents: 3432
diff changeset
2621
7dc8e4a12105 New functions in DSPContext for VC-1 decoding
kostya
parents: 3432
diff changeset
2622 void ff_put_vc1_mspel_mc00_c(uint8_t *dst, uint8_t *src, int stride, int rnd) {
7dc8e4a12105 New functions in DSPContext for VC-1 decoding
kostya
parents: 3432
diff changeset
2623 put_pixels8_c(dst, src, stride, 8);
7dc8e4a12105 New functions in DSPContext for VC-1 decoding
kostya
parents: 3432
diff changeset
2624 }
3656
1cc5bdadd487 New qpel MC functions conforming to VC-1 standard.
kostya
parents: 3645
diff changeset
2625
1cc5bdadd487 New qpel MC functions conforming to VC-1 standard.
kostya
parents: 3645
diff changeset
2626 void ff_put_vc1_qpel_mc00_c(uint8_t *dst, uint8_t *src, int stride, int rnd) {
1cc5bdadd487 New qpel MC functions conforming to VC-1 standard.
kostya
parents: 3645
diff changeset
2627 put_pixels8_c(dst, src, stride, 8);
1cc5bdadd487 New qpel MC functions conforming to VC-1 standard.
kostya
parents: 3645
diff changeset
2628 }
3526
7dc8e4a12105 New functions in DSPContext for VC-1 decoding
kostya
parents: 3432
diff changeset
2629 #endif /* CONFIG_VC1_DECODER||CONFIG_WMV3_DECODER */
7dc8e4a12105 New functions in DSPContext for VC-1 decoding
kostya
parents: 3432
diff changeset
2630
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2631 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
2632 uint8_t *cm = cropTbl + MAX_NEG_CROP;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2633 int i;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2634
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2635 for(i=0; i<w; i++){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2636 const int src_1= src[ -srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2637 const int src0 = src[0 ];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2638 const int src1 = src[ srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2639 const int src2 = src[2*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2640 const int src3 = src[3*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2641 const int src4 = src[4*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2642 const int src5 = src[5*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2643 const int src6 = src[6*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2644 const int src7 = src[7*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2645 const int src8 = src[8*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2646 const int src9 = src[9*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2647 dst[0*dstStride]= cm[(9*(src0 + src1) - (src_1 + src2) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2648 dst[1*dstStride]= cm[(9*(src1 + src2) - (src0 + src3) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2649 dst[2*dstStride]= cm[(9*(src2 + src3) - (src1 + src4) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2650 dst[3*dstStride]= cm[(9*(src3 + src4) - (src2 + src5) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2651 dst[4*dstStride]= cm[(9*(src4 + src5) - (src3 + src6) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2652 dst[5*dstStride]= cm[(9*(src5 + src6) - (src4 + src7) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2653 dst[6*dstStride]= cm[(9*(src6 + src7) - (src5 + src8) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2654 dst[7*dstStride]= cm[(9*(src7 + src8) - (src6 + src9) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2655 src++;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2656 dst++;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2657 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2658 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2659
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2660 static void put_mspel8_mc00_c (uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2661 put_pixels8_c(dst, src, stride, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2662 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2663
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2664 static void put_mspel8_mc10_c(uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2665 uint8_t half[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2666 wmv2_mspel8_h_lowpass(half, src, 8, stride, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2667 put_pixels8_l2(dst, src, half, stride, stride, 8, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2668 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2669
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2670 static void put_mspel8_mc20_c(uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2671 wmv2_mspel8_h_lowpass(dst, src, stride, stride, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2672 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2673
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2674 static void put_mspel8_mc30_c(uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2675 uint8_t half[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2676 wmv2_mspel8_h_lowpass(half, src, 8, stride, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2677 put_pixels8_l2(dst, src+1, half, stride, stride, 8, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2678 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2679
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2680 static void put_mspel8_mc02_c(uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2681 wmv2_mspel8_v_lowpass(dst, src, stride, stride, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2682 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2683
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2684 static void put_mspel8_mc12_c(uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2685 uint8_t halfH[88];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2686 uint8_t halfV[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2687 uint8_t halfHV[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2688 wmv2_mspel8_h_lowpass(halfH, src-stride, 8, stride, 11);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2689 wmv2_mspel8_v_lowpass(halfV, src, 8, stride, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2690 wmv2_mspel8_v_lowpass(halfHV, halfH+8, 8, 8, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2691 put_pixels8_l2(dst, halfV, halfHV, stride, 8, 8, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2692 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2693 static void put_mspel8_mc32_c(uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2694 uint8_t halfH[88];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2695 uint8_t halfV[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2696 uint8_t halfHV[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2697 wmv2_mspel8_h_lowpass(halfH, src-stride, 8, stride, 11);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2698 wmv2_mspel8_v_lowpass(halfV, src+1, 8, stride, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2699 wmv2_mspel8_v_lowpass(halfHV, halfH+8, 8, 8, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2700 put_pixels8_l2(dst, halfV, halfHV, stride, 8, 8, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2701 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2702 static void put_mspel8_mc22_c(uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2703 uint8_t halfH[88];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2704 wmv2_mspel8_h_lowpass(halfH, src-stride, 8, stride, 11);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2705 wmv2_mspel8_v_lowpass(dst, halfH+8, stride, 8, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2706 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2707
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2708 static void h263_v_loop_filter_c(uint8_t *src, int stride, int qscale){
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2709 int x;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2710 const int strength= ff_h263_loop_filter_strength[qscale];
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2711
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2712 for(x=0; x<8; x++){
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2713 int d1, d2, ad1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2714 int p0= src[x-2*stride];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2715 int p1= src[x-1*stride];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2716 int p2= src[x+0*stride];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2717 int p3= src[x+1*stride];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2718 int d = (p0 - p3 + 4*(p2 - p1)) / 8;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2719
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2720 if (d<-2*strength) d1= 0;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2721 else if(d<- strength) d1=-2*strength - d;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2722 else if(d< strength) d1= d;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2723 else if(d< 2*strength) d1= 2*strength - d;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2724 else d1= 0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2725
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2726 p1 += d1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2727 p2 -= d1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2728 if(p1&256) p1= ~(p1>>31);
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2729 if(p2&256) p2= ~(p2>>31);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2730
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2731 src[x-1*stride] = p1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2732 src[x+0*stride] = p2;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2733
1645
michael
parents: 1644
diff changeset
2734 ad1= ABS(d1)>>1;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2735
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2736 d2= clip((p0-p3)/4, -ad1, ad1);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2737
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2738 src[x-2*stride] = p0 - d2;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2739 src[x+ stride] = p3 + d2;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2740 }
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2741 }
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2742
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2743 static void h263_h_loop_filter_c(uint8_t *src, int stride, int qscale){
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2744 int y;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2745 const int strength= ff_h263_loop_filter_strength[qscale];
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2746
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2747 for(y=0; y<8; y++){
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2748 int d1, d2, ad1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2749 int p0= src[y*stride-2];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2750 int p1= src[y*stride-1];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2751 int p2= src[y*stride+0];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2752 int p3= src[y*stride+1];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2753 int d = (p0 - p3 + 4*(p2 - p1)) / 8;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2754
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2755 if (d<-2*strength) d1= 0;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2756 else if(d<- strength) d1=-2*strength - d;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2757 else if(d< strength) d1= d;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2758 else if(d< 2*strength) d1= 2*strength - d;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2759 else d1= 0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2760
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2761 p1 += d1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2762 p2 -= d1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2763 if(p1&256) p1= ~(p1>>31);
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2764 if(p2&256) p2= ~(p2>>31);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2765
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2766 src[y*stride-1] = p1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2767 src[y*stride+0] = p2;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2768
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2769 ad1= ABS(d1)>>1;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2770
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2771 d2= clip((p0-p3)/4, -ad1, ad1);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2772
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2773 src[y*stride-2] = p0 - d2;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2774 src[y*stride+1] = p3 + d2;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2775 }
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2776 }
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2777
2045
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2778 static void h261_loop_filter_c(uint8_t *src, int stride){
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2779 int x,y,xy,yz;
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2780 int temp[64];
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2781
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2782 for(x=0; x<8; x++){
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2783 temp[x ] = 4*src[x ];
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2784 temp[x + 7*8] = 4*src[x + 7*stride];
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2785 }
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2786 for(y=1; y<7; y++){
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2787 for(x=0; x<8; x++){
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2788 xy = y * stride + x;
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2789 yz = y * 8 + x;
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2790 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
2791 }
b6f2add2511e h261 decoder by (Maarten Daniels <maarten.daniels at student dot luc dot ac dot be>)
michael
parents: 1984
diff changeset
2792 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2793
2045
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2794 for(y=0; y<8; y++){
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2795 src[ y*stride] = (temp[ y*8] + 2)>>2;
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2796 src[7+y*stride] = (temp[7+y*8] + 2)>>2;
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2797 for(x=1; x<7; x++){
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2798 xy = y * stride + x;
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2799 yz = y * 8 + x;
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2800 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
2801 }
b6f2add2511e h261 decoder by (Maarten Daniels <maarten.daniels at student dot luc dot ac dot be>)
michael
parents: 1984
diff changeset
2802 }
b6f2add2511e h261 decoder by (Maarten Daniels <maarten.daniels at student dot luc dot ac dot be>)
michael
parents: 1984
diff changeset
2803 }
b6f2add2511e h261 decoder by (Maarten Daniels <maarten.daniels at student dot luc dot ac dot be>)
michael
parents: 1984
diff changeset
2804
2707
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2805 static inline void h264_loop_filter_luma_c(uint8_t *pix, int xstride, int ystride, int alpha, int beta, int8_t *tc0)
2633
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2806 {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2807 int i, d;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2808 for( i = 0; i < 4; i++ ) {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2809 if( tc0[i] < 0 ) {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2810 pix += 4*ystride;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2811 continue;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2812 }
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2813 for( d = 0; d < 4; d++ ) {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2814 const int p0 = pix[-1*xstride];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2815 const int p1 = pix[-2*xstride];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2816 const int p2 = pix[-3*xstride];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2817 const int q0 = pix[0];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2818 const int q1 = pix[1*xstride];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2819 const int q2 = pix[2*xstride];
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2820
2633
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2821 if( ABS( p0 - q0 ) < alpha &&
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2822 ABS( p1 - p0 ) < beta &&
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2823 ABS( q1 - q0 ) < beta ) {
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2824
2633
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2825 int tc = tc0[i];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2826 int i_delta;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2827
2633
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2828 if( ABS( p2 - p0 ) < beta ) {
2651
534f3b47afda minor optimization of the h264 loop filter
michael
parents: 2639
diff changeset
2829 pix[-2*xstride] = p1 + clip( (( p2 + ( ( p0 + q0 + 1 ) >> 1 ) ) >> 1) - p1, -tc0[i], tc0[i] );
2633
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2830 tc++;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2831 }
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2832 if( ABS( q2 - q0 ) < beta ) {
2651
534f3b47afda minor optimization of the h264 loop filter
michael
parents: 2639
diff changeset
2833 pix[ xstride] = q1 + clip( (( q2 + ( ( p0 + q0 + 1 ) >> 1 ) ) >> 1) - q1, -tc0[i], tc0[i] );
2633
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2834 tc++;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2835 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2836
2633
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2837 i_delta = clip( (((q0 - p0 ) << 2) + (p1 - q1) + 4) >> 3, -tc, tc );
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2838 pix[-xstride] = clip_uint8( p0 + i_delta ); /* p0' */
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2839 pix[0] = clip_uint8( q0 - i_delta ); /* q0' */
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2840 }
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2841 pix += ystride;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2842 }
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2843 }
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2844 }
2707
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2845 static void h264_v_loop_filter_luma_c(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0)
2633
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2846 {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2847 h264_loop_filter_luma_c(pix, stride, 1, alpha, beta, tc0);
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2848 }
2707
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2849 static void h264_h_loop_filter_luma_c(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0)
2633
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2850 {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2851 h264_loop_filter_luma_c(pix, 1, stride, alpha, beta, tc0);
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2852 }
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2853
2707
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2854 static inline void h264_loop_filter_chroma_c(uint8_t *pix, int xstride, int ystride, int alpha, int beta, int8_t *tc0)
2633
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2855 {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2856 int i, d;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2857 for( i = 0; i < 4; i++ ) {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2858 const int tc = tc0[i];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2859 if( tc <= 0 ) {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2860 pix += 2*ystride;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2861 continue;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2862 }
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2863 for( d = 0; d < 2; d++ ) {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2864 const int p0 = pix[-1*xstride];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2865 const int p1 = pix[-2*xstride];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2866 const int q0 = pix[0];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2867 const int q1 = pix[1*xstride];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2868
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2869 if( ABS( p0 - q0 ) < alpha &&
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2870 ABS( p1 - p0 ) < beta &&
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2871 ABS( q1 - q0 ) < beta ) {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2872
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2873 int delta = clip( (((q0 - p0 ) << 2) + (p1 - q1) + 4) >> 3, -tc, tc );
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2874
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2875 pix[-xstride] = clip_uint8( p0 + delta ); /* p0' */
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2876 pix[0] = clip_uint8( q0 - delta ); /* q0' */
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2877 }
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2878 pix += ystride;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2879 }
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2880 }
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2881 }
2707
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2882 static void h264_v_loop_filter_chroma_c(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0)
2633
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2883 {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2884 h264_loop_filter_chroma_c(pix, stride, 1, alpha, beta, tc0);
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2885 }
2707
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2886 static void h264_h_loop_filter_chroma_c(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0)
2633
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2887 {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2888 h264_loop_filter_chroma_c(pix, 1, stride, alpha, beta, tc0);
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2889 }
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
2890
2707
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2891 static inline void h264_loop_filter_chroma_intra_c(uint8_t *pix, int xstride, int ystride, int alpha, int beta)
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2892 {
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2893 int d;
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2894 for( d = 0; d < 8; d++ ) {
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2895 const int p0 = pix[-1*xstride];
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2896 const int p1 = pix[-2*xstride];
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2897 const int q0 = pix[0];
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2898 const int q1 = pix[1*xstride];
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2899
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2900 if( ABS( p0 - q0 ) < alpha &&
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2901 ABS( p1 - p0 ) < beta &&
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2902 ABS( q1 - q0 ) < beta ) {
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2903
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2904 pix[-xstride] = ( 2*p1 + p0 + q1 + 2 ) >> 2; /* p0' */
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2905 pix[0] = ( 2*q1 + q0 + p1 + 2 ) >> 2; /* q0' */
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2906 }
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2907 pix += ystride;
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2908 }
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2909 }
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2910 static void h264_v_loop_filter_chroma_intra_c(uint8_t *pix, int stride, int alpha, int beta)
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2911 {
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2912 h264_loop_filter_chroma_intra_c(pix, stride, 1, alpha, beta);
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2913 }
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2914 static void h264_h_loop_filter_chroma_intra_c(uint8_t *pix, int stride, int alpha, int beta)
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2915 {
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2916 h264_loop_filter_chroma_intra_c(pix, 1, stride, alpha, beta);
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2917 }
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
2918
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2919 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
2920 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
2921 int s, i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2922
986e461dc072 Initial revision
glantau
parents:
diff changeset
2923 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2924 for(i=0;i<h;i++) {
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
2925 s += abs(pix1[0] - pix2[0]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2926 s += abs(pix1[1] - pix2[1]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2927 s += abs(pix1[2] - pix2[2]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2928 s += abs(pix1[3] - pix2[3]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2929 s += abs(pix1[4] - pix2[4]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2930 s += abs(pix1[5] - pix2[5]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2931 s += abs(pix1[6] - pix2[6]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2932 s += abs(pix1[7] - pix2[7]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2933 s += abs(pix1[8] - pix2[8]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2934 s += abs(pix1[9] - pix2[9]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2935 s += abs(pix1[10] - pix2[10]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2936 s += abs(pix1[11] - pix2[11]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2937 s += abs(pix1[12] - pix2[12]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2938 s += abs(pix1[13] - pix2[13]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2939 s += abs(pix1[14] - pix2[14]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2940 s += abs(pix1[15] - pix2[15]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
2941 pix1 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2942 pix2 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2943 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
2944 return s;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2945 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
2946
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2947 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
2948 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
2949 int s, i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2950
986e461dc072 Initial revision
glantau
parents:
diff changeset
2951 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2952 for(i=0;i<h;i++) {
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
2953 s += abs(pix1[0] - avg2(pix2[0], pix2[1]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2954 s += abs(pix1[1] - avg2(pix2[1], pix2[2]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2955 s += abs(pix1[2] - avg2(pix2[2], pix2[3]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2956 s += abs(pix1[3] - avg2(pix2[3], pix2[4]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2957 s += abs(pix1[4] - avg2(pix2[4], pix2[5]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2958 s += abs(pix1[5] - avg2(pix2[5], pix2[6]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2959 s += abs(pix1[6] - avg2(pix2[6], pix2[7]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2960 s += abs(pix1[7] - avg2(pix2[7], pix2[8]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2961 s += abs(pix1[8] - avg2(pix2[8], pix2[9]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2962 s += abs(pix1[9] - avg2(pix2[9], pix2[10]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2963 s += abs(pix1[10] - avg2(pix2[10], pix2[11]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2964 s += abs(pix1[11] - avg2(pix2[11], pix2[12]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2965 s += abs(pix1[12] - avg2(pix2[12], pix2[13]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2966 s += abs(pix1[13] - avg2(pix2[13], pix2[14]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2967 s += abs(pix1[14] - avg2(pix2[14], pix2[15]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2968 s += abs(pix1[15] - avg2(pix2[15], pix2[16]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2969 pix1 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2970 pix2 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2971 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
2972 return s;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2973 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
2974
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2975 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
2976 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
2977 int s, i;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2978 uint8_t *pix3 = pix2 + line_size;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
2979
986e461dc072 Initial revision
glantau
parents:
diff changeset
2980 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
2981 for(i=0;i<h;i++) {
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
2982 s += abs(pix1[0] - avg2(pix2[0], pix3[0]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2983 s += abs(pix1[1] - avg2(pix2[1], pix3[1]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2984 s += abs(pix1[2] - avg2(pix2[2], pix3[2]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2985 s += abs(pix1[3] - avg2(pix2[3], pix3[3]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2986 s += abs(pix1[4] - avg2(pix2[4], pix3[4]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2987 s += abs(pix1[5] - avg2(pix2[5], pix3[5]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2988 s += abs(pix1[6] - avg2(pix2[6], pix3[6]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2989 s += abs(pix1[7] - avg2(pix2[7], pix3[7]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2990 s += abs(pix1[8] - avg2(pix2[8], pix3[8]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2991 s += abs(pix1[9] - avg2(pix2[9], pix3[9]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2992 s += abs(pix1[10] - avg2(pix2[10], pix3[10]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2993 s += abs(pix1[11] - avg2(pix2[11], pix3[11]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2994 s += abs(pix1[12] - avg2(pix2[12], pix3[12]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2995 s += abs(pix1[13] - avg2(pix2[13], pix3[13]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2996 s += abs(pix1[14] - avg2(pix2[14], pix3[14]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2997 s += abs(pix1[15] - avg2(pix2[15], pix3[15]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
2998 pix1 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
2999 pix2 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3000 pix3 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3001 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
3002 return s;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3003 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
3004
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3005 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
3006 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
3007 int s, i;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
3008 uint8_t *pix3 = pix2 + line_size;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
3009
986e461dc072 Initial revision
glantau
parents:
diff changeset
3010 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3011 for(i=0;i<h;i++) {
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
3012 s += abs(pix1[0] - avg4(pix2[0], pix2[1], pix3[0], pix3[1]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3013 s += abs(pix1[1] - avg4(pix2[1], pix2[2], pix3[1], pix3[2]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3014 s += abs(pix1[2] - avg4(pix2[2], pix2[3], pix3[2], pix3[3]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3015 s += abs(pix1[3] - avg4(pix2[3], pix2[4], pix3[3], pix3[4]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3016 s += abs(pix1[4] - avg4(pix2[4], pix2[5], pix3[4], pix3[5]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3017 s += abs(pix1[5] - avg4(pix2[5], pix2[6], pix3[5], pix3[6]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3018 s += abs(pix1[6] - avg4(pix2[6], pix2[7], pix3[6], pix3[7]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3019 s += abs(pix1[7] - avg4(pix2[7], pix2[8], pix3[7], pix3[8]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3020 s += abs(pix1[8] - avg4(pix2[8], pix2[9], pix3[8], pix3[9]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3021 s += abs(pix1[9] - avg4(pix2[9], pix2[10], pix3[9], pix3[10]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3022 s += abs(pix1[10] - avg4(pix2[10], pix2[11], pix3[10], pix3[11]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3023 s += abs(pix1[11] - avg4(pix2[11], pix2[12], pix3[11], pix3[12]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3024 s += abs(pix1[12] - avg4(pix2[12], pix2[13], pix3[12], pix3[13]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3025 s += abs(pix1[13] - avg4(pix2[13], pix2[14], pix3[13], pix3[14]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3026 s += abs(pix1[14] - avg4(pix2[14], pix2[15], pix3[14], pix3[15]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3027 s += abs(pix1[15] - avg4(pix2[15], pix2[16], pix3[15], pix3[16]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3028 pix1 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3029 pix2 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3030 pix3 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3031 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
3032 return s;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3033 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
3034
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3035 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
3036 {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3037 int s, i;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3038
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3039 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3040 for(i=0;i<h;i++) {
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3041 s += abs(pix1[0] - pix2[0]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3042 s += abs(pix1[1] - pix2[1]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3043 s += abs(pix1[2] - pix2[2]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3044 s += abs(pix1[3] - pix2[3]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3045 s += abs(pix1[4] - pix2[4]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3046 s += abs(pix1[5] - pix2[5]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3047 s += abs(pix1[6] - pix2[6]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3048 s += abs(pix1[7] - pix2[7]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3049 pix1 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3050 pix2 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3051 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3052 return s;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3053 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3054
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3055 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
3056 {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3057 int s, i;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3058
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3059 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3060 for(i=0;i<h;i++) {
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3061 s += abs(pix1[0] - avg2(pix2[0], pix2[1]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3062 s += abs(pix1[1] - avg2(pix2[1], pix2[2]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3063 s += abs(pix1[2] - avg2(pix2[2], pix2[3]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3064 s += abs(pix1[3] - avg2(pix2[3], pix2[4]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3065 s += abs(pix1[4] - avg2(pix2[4], pix2[5]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3066 s += abs(pix1[5] - avg2(pix2[5], pix2[6]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3067 s += abs(pix1[6] - avg2(pix2[6], pix2[7]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3068 s += abs(pix1[7] - avg2(pix2[7], pix2[8]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3069 pix1 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3070 pix2 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3071 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3072 return s;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3073 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3074
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3075 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
3076 {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3077 int s, i;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
3078 uint8_t *pix3 = pix2 + line_size;
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3079
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3080 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3081 for(i=0;i<h;i++) {
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3082 s += abs(pix1[0] - avg2(pix2[0], pix3[0]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3083 s += abs(pix1[1] - avg2(pix2[1], pix3[1]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3084 s += abs(pix1[2] - avg2(pix2[2], pix3[2]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3085 s += abs(pix1[3] - avg2(pix2[3], pix3[3]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3086 s += abs(pix1[4] - avg2(pix2[4], pix3[4]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3087 s += abs(pix1[5] - avg2(pix2[5], pix3[5]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3088 s += abs(pix1[6] - avg2(pix2[6], pix3[6]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3089 s += abs(pix1[7] - avg2(pix2[7], pix3[7]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3090 pix1 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3091 pix2 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3092 pix3 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3093 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3094 return s;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3095 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3096
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3097 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
3098 {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3099 int s, i;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
3100 uint8_t *pix3 = pix2 + line_size;
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3101
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3102 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3103 for(i=0;i<h;i++) {
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3104 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
3105 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
3106 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
3107 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
3108 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
3109 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
3110 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
3111 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
3112 pix1 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3113 pix2 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3114 pix3 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3115 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3116 return s;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3117 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3118
2834
fd5d7c732c6b kill a bunch of compiler warnings
mru
parents: 2763
diff changeset
3119 static int nsse16_c(void *v, uint8_t *s1, uint8_t *s2, int stride, int h){
fd5d7c732c6b kill a bunch of compiler warnings
mru
parents: 2763
diff changeset
3120 MpegEncContext *c = v;
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3121 int score1=0;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3122 int score2=0;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3123 int x,y;
2066
4bfb146e701b nsse weight
michael
parents: 2065
diff changeset
3124
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3125 for(y=0; y<h; y++){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3126 for(x=0; x<16; x++){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3127 score1+= (s1[x ] - s2[x ])*(s1[x ] - s2[x ]);
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3128 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3129 if(y+1<h){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3130 for(x=0; x<15; x++){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3131 score2+= ABS( s1[x ] - s1[x +stride]
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3132 - s1[x+1] + s1[x+1+stride])
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3133 -ABS( s2[x ] - s2[x +stride]
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3134 - s2[x+1] + s2[x+1+stride]);
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3135 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3136 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3137 s1+= stride;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3138 s2+= stride;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3139 }
2066
4bfb146e701b nsse weight
michael
parents: 2065
diff changeset
3140
4bfb146e701b nsse weight
michael
parents: 2065
diff changeset
3141 if(c) return score1 + ABS(score2)*c->avctx->nsse_weight;
4bfb146e701b nsse weight
michael
parents: 2065
diff changeset
3142 else return score1 + ABS(score2)*8;
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3143 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3144
2834
fd5d7c732c6b kill a bunch of compiler warnings
mru
parents: 2763
diff changeset
3145 static int nsse8_c(void *v, uint8_t *s1, uint8_t *s2, int stride, int h){
fd5d7c732c6b kill a bunch of compiler warnings
mru
parents: 2763
diff changeset
3146 MpegEncContext *c = v;
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3147 int score1=0;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3148 int score2=0;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3149 int x,y;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3150
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3151 for(y=0; y<h; y++){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3152 for(x=0; x<8; x++){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3153 score1+= (s1[x ] - s2[x ])*(s1[x ] - s2[x ]);
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3154 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3155 if(y+1<h){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3156 for(x=0; x<7; x++){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3157 score2+= ABS( s1[x ] - s1[x +stride]
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3158 - s1[x+1] + s1[x+1+stride])
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3159 -ABS( s2[x ] - s2[x +stride]
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3160 - s2[x+1] + s2[x+1+stride]);
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3161 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3162 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3163 s1+= stride;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3164 s2+= stride;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3165 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3166
2066
4bfb146e701b nsse weight
michael
parents: 2065
diff changeset
3167 if(c) return score1 + ABS(score2)*c->avctx->nsse_weight;
4bfb146e701b nsse weight
michael
parents: 2065
diff changeset
3168 else return score1 + ABS(score2)*8;
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3169 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3170
1784
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3171 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
3172 int i;
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3173 unsigned int sum=0;
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3174
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3175 for(i=0; i<8*8; i++){
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3176 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
3177 int w= weight[i];
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3178 b>>= RECON_SHIFT;
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3179 assert(-512<b && b<512);
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3180
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3181 sum += (w*b)*(w*b)>>4;
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3182 }
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3183 return sum>>2;
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3184 }
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3185
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3186 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
3187 int i;
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3188
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3189 for(i=0; i<8*8; i++){
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3190 rem[i] += (basis[i]*scale + (1<<(BASIS_SHIFT - RECON_SHIFT-1)))>>(BASIS_SHIFT - RECON_SHIFT);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3191 }
1784
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3192 }
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3193
1100
207f27932628 ff_block_permute() dox
michaelni
parents: 1092
diff changeset
3194 /**
207f27932628 ff_block_permute() dox
michaelni
parents: 1092
diff changeset
3195 * permutes an 8x8 block.
1101
a3a8017dee70 more idiot proof dox
michaelni
parents: 1100
diff changeset
3196 * @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
3197 * @param permutation the permutation vector
207f27932628 ff_block_permute() dox
michaelni
parents: 1092
diff changeset
3198 * @param last the last non zero coefficient in scantable order, used to speed the permutation up
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3199 * @param scantable the used scantable, this is only used to speed the permutation up, the block is not
1101
a3a8017dee70 more idiot proof dox
michaelni
parents: 1100
diff changeset
3200 * (inverse) permutated to scantable order!
1100
207f27932628 ff_block_permute() dox
michaelni
parents: 1092
diff changeset
3201 */
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
3202 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
3203 {
764
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3204 int i;
945
463f7260b155 trellis quantization
michaelni
parents: 936
diff changeset
3205 DCTELEM temp[64];
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3206
764
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3207 if(last<=0) return;
882
34943cb0c2fb * remaining part of John Ryland's patch
kabi
parents: 866
diff changeset
3208 //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
3209
764
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3210 for(i=0; i<=last; i++){
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3211 const int j= scantable[i];
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3212 temp[j]= block[j];
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3213 block[j]=0;
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3214 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3215
764
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3216 for(i=0; i<=last; i++){
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3217 const int j= scantable[i];
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3218 const int perm_j= permutation[j];
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3219 block[perm_j]= temp[j];
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3220 }
174
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
3221 }
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
3222
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3223 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
3224 return 0;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3225 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3226
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3227 void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3228 int i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3229
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3230 memset(cmp, 0, sizeof(void*)*5);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3231
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3232 for(i=0; i<5; i++){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3233 switch(type&0xFF){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3234 case FF_CMP_SAD:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3235 cmp[i]= c->sad[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3236 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3237 case FF_CMP_SATD:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3238 cmp[i]= c->hadamard8_diff[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3239 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3240 case FF_CMP_SSE:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3241 cmp[i]= c->sse[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3242 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3243 case FF_CMP_DCT:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3244 cmp[i]= c->dct_sad[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3245 break;
3010
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3246 case FF_CMP_DCT264:
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3247 cmp[i]= c->dct264_sad[i];
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3248 break;
2382
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3249 case FF_CMP_DCTMAX:
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3250 cmp[i]= c->dct_max[i];
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3251 break;
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3252 case FF_CMP_PSNR:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3253 cmp[i]= c->quant_psnr[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3254 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3255 case FF_CMP_BIT:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3256 cmp[i]= c->bit[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3257 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3258 case FF_CMP_RD:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3259 cmp[i]= c->rd[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3260 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3261 case FF_CMP_VSAD:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3262 cmp[i]= c->vsad[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3263 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3264 case FF_CMP_VSSE:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3265 cmp[i]= c->vsse[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3266 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3267 case FF_CMP_ZERO:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3268 cmp[i]= zero_cmp;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3269 break;
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3270 case FF_CMP_NSSE:
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3271 cmp[i]= c->nsse[i];
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3272 break;
3373
b8996cc5ccae Disable w53 and w97 cmp methods when snow encoder is disabled
gpoirier
parents: 3323
diff changeset
3273 #ifdef CONFIG_SNOW_ENCODER
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
3274 case FF_CMP_W53:
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
3275 cmp[i]= c->w53[i];
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
3276 break;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
3277 case FF_CMP_W97:
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
3278 cmp[i]= c->w97[i];
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
3279 break;
3373
b8996cc5ccae Disable w53 and w97 cmp methods when snow encoder is disabled
gpoirier
parents: 3323
diff changeset
3280 #endif
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3281 default:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3282 av_log(NULL, AV_LOG_ERROR,"internal error in cmp function selection\n");
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3283 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3284 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3285 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3286
1101
a3a8017dee70 more idiot proof dox
michaelni
parents: 1100
diff changeset
3287 /**
a3a8017dee70 more idiot proof dox
michaelni
parents: 1100
diff changeset
3288 * memset(blocks, 0, sizeof(DCTELEM)*6*64)
a3a8017dee70 more idiot proof dox
michaelni
parents: 1100
diff changeset
3289 */
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
3290 static void clear_blocks_c(DCTELEM *blocks)
296
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
3291 {
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
3292 memset(blocks, 0, sizeof(DCTELEM)*6*64);
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
3293 }
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
3294
866
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3295 static void add_bytes_c(uint8_t *dst, uint8_t *src, int w){
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3296 int i;
996
ad44196ea483 add/diff_bytes bugfix patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>)
michaelni
parents: 984
diff changeset
3297 for(i=0; i+7<w; i+=8){
866
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3298 dst[i+0] += src[i+0];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3299 dst[i+1] += src[i+1];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3300 dst[i+2] += src[i+2];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3301 dst[i+3] += src[i+3];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3302 dst[i+4] += src[i+4];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3303 dst[i+5] += src[i+5];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3304 dst[i+6] += src[i+6];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3305 dst[i+7] += src[i+7];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3306 }
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3307 for(; i<w; i++)
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3308 dst[i+0] += src[i+0];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3309 }
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3310
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3311 static void diff_bytes_c(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w){
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3312 int i;
996
ad44196ea483 add/diff_bytes bugfix patch by (Felix von Leitner <felix-ffmpeg at fefe dot de>)
michaelni
parents: 984
diff changeset
3313 for(i=0; i+7<w; i+=8){
866
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3314 dst[i+0] = src1[i+0]-src2[i+0];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3315 dst[i+1] = src1[i+1]-src2[i+1];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3316 dst[i+2] = src1[i+2]-src2[i+2];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3317 dst[i+3] = src1[i+3]-src2[i+3];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3318 dst[i+4] = src1[i+4]-src2[i+4];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3319 dst[i+5] = src1[i+5]-src2[i+5];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3320 dst[i+6] = src1[i+6]-src2[i+6];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3321 dst[i+7] = src1[i+7]-src2[i+7];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3322 }
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3323 for(; i<w; i++)
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3324 dst[i+0] = src1[i+0]-src2[i+0];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3325 }
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3326
1527
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3327 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
3328 int i;
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3329 uint8_t l, lt;
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3330
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3331 l= *left;
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3332 lt= *left_top;
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3333
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3334 for(i=0; i<w; i++){
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3335 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
3336 lt= src1[i];
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3337 l= src2[i];
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3338 dst[i]= l - pred;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3339 }
1527
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3340
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3341 *left= l;
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3342 *left_top= lt;
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3343 }
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3344
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3345 #define BUTTERFLY2(o1,o2,i1,i2) \
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3346 o1= (i1)+(i2);\
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3347 o2= (i1)-(i2);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3348
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3349 #define BUTTERFLY1(x,y) \
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3350 {\
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3351 int a,b;\
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3352 a= x;\
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3353 b= y;\
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3354 x= a+b;\
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3355 y= a-b;\
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3356 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3357
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3358 #define BUTTERFLYA(x,y) (ABS((x)+(y)) + ABS((x)-(y)))
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3359
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3360 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
3361 int i;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3362 int temp[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3363 int sum=0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3364
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3365 assert(h==8);
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3366
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3367 for(i=0; i<8; i++){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3368 //FIXME try pointer walks
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3369 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
3370 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
3371 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
3372 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]);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3373
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3374 BUTTERFLY1(temp[8*i+0], temp[8*i+2]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3375 BUTTERFLY1(temp[8*i+1], temp[8*i+3]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3376 BUTTERFLY1(temp[8*i+4], temp[8*i+6]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3377 BUTTERFLY1(temp[8*i+5], temp[8*i+7]);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3378
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3379 BUTTERFLY1(temp[8*i+0], temp[8*i+4]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3380 BUTTERFLY1(temp[8*i+1], temp[8*i+5]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3381 BUTTERFLY1(temp[8*i+2], temp[8*i+6]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3382 BUTTERFLY1(temp[8*i+3], temp[8*i+7]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3383 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3384
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3385 for(i=0; i<8; i++){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3386 BUTTERFLY1(temp[8*0+i], temp[8*1+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3387 BUTTERFLY1(temp[8*2+i], temp[8*3+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3388 BUTTERFLY1(temp[8*4+i], temp[8*5+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3389 BUTTERFLY1(temp[8*6+i], temp[8*7+i]);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3390
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3391 BUTTERFLY1(temp[8*0+i], temp[8*2+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3392 BUTTERFLY1(temp[8*1+i], temp[8*3+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3393 BUTTERFLY1(temp[8*4+i], temp[8*6+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3394 BUTTERFLY1(temp[8*5+i], temp[8*7+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3395
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3396 sum +=
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3397 BUTTERFLYA(temp[8*0+i], temp[8*4+i])
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3398 +BUTTERFLYA(temp[8*1+i], temp[8*5+i])
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3399 +BUTTERFLYA(temp[8*2+i], temp[8*6+i])
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3400 +BUTTERFLYA(temp[8*3+i], temp[8*7+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3401 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3402 #if 0
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3403 static int maxi=0;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3404 if(sum>maxi){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3405 maxi=sum;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3406 printf("MAX:%d\n", maxi);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3407 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3408 #endif
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3409 return sum;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3410 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3411
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3412 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
3413 int i;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3414 int temp[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3415 int sum=0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3416
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3417 assert(h==8);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3418
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3419 for(i=0; i<8; i++){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3420 //FIXME try pointer walks
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3421 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
3422 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
3423 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
3424 BUTTERFLY2(temp[8*i+6], temp[8*i+7], src[stride*i+6],src[stride*i+7]);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3425
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3426 BUTTERFLY1(temp[8*i+0], temp[8*i+2]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3427 BUTTERFLY1(temp[8*i+1], temp[8*i+3]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3428 BUTTERFLY1(temp[8*i+4], temp[8*i+6]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3429 BUTTERFLY1(temp[8*i+5], temp[8*i+7]);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3430
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3431 BUTTERFLY1(temp[8*i+0], temp[8*i+4]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3432 BUTTERFLY1(temp[8*i+1], temp[8*i+5]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3433 BUTTERFLY1(temp[8*i+2], temp[8*i+6]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3434 BUTTERFLY1(temp[8*i+3], temp[8*i+7]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3435 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3436
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3437 for(i=0; i<8; i++){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3438 BUTTERFLY1(temp[8*0+i], temp[8*1+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3439 BUTTERFLY1(temp[8*2+i], temp[8*3+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3440 BUTTERFLY1(temp[8*4+i], temp[8*5+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3441 BUTTERFLY1(temp[8*6+i], temp[8*7+i]);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3442
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3443 BUTTERFLY1(temp[8*0+i], temp[8*2+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3444 BUTTERFLY1(temp[8*1+i], temp[8*3+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3445 BUTTERFLY1(temp[8*4+i], temp[8*6+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3446 BUTTERFLY1(temp[8*5+i], temp[8*7+i]);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3447
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3448 sum +=
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3449 BUTTERFLYA(temp[8*0+i], temp[8*4+i])
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3450 +BUTTERFLYA(temp[8*1+i], temp[8*5+i])
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3451 +BUTTERFLYA(temp[8*2+i], temp[8*6+i])
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3452 +BUTTERFLYA(temp[8*3+i], temp[8*7+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3453 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3454
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3455 sum -= ABS(temp[8*0] + temp[8*4]); // -mean
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3456
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3457 return sum;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3458 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3459
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3460 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
3461 MpegEncContext * const s= (MpegEncContext *)c;
3089
072dbc669253 MSVC-compatible __align8/__align16 declaration
diego
parents: 3036
diff changeset
3462 DECLARE_ALIGNED_8(uint64_t, aligned_temp[sizeof(DCTELEM)*64/8]);
1016
michaelni
parents: 1013
diff changeset
3463 DCTELEM * const temp= (DCTELEM*)aligned_temp;
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3464 int sum=0, i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3465
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3466 assert(h==8);
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3467
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3468 s->dsp.diff_pixels(temp, src1, src2, stride);
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3469 s->dsp.fdct(temp);
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3470
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3471 for(i=0; i<64; i++)
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3472 sum+= ABS(temp[i]);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3473
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3474 return sum;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3475 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3476
3010
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3477 #ifdef CONFIG_GPL
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3478 #define DCT8_1D {\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3479 const int s07 = SRC(0) + SRC(7);\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3480 const int s16 = SRC(1) + SRC(6);\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3481 const int s25 = SRC(2) + SRC(5);\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3482 const int s34 = SRC(3) + SRC(4);\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3483 const int a0 = s07 + s34;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3484 const int a1 = s16 + s25;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3485 const int a2 = s07 - s34;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3486 const int a3 = s16 - s25;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3487 const int d07 = SRC(0) - SRC(7);\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3488 const int d16 = SRC(1) - SRC(6);\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3489 const int d25 = SRC(2) - SRC(5);\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3490 const int d34 = SRC(3) - SRC(4);\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3491 const int a4 = d16 + d25 + (d07 + (d07>>1));\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3492 const int a5 = d07 - d34 - (d25 + (d25>>1));\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3493 const int a6 = d07 + d34 - (d16 + (d16>>1));\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3494 const int a7 = d16 - d25 + (d34 + (d34>>1));\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3495 DST(0, a0 + a1 ) ;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3496 DST(1, a4 + (a7>>2)) ;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3497 DST(2, a2 + (a3>>1)) ;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3498 DST(3, a5 + (a6>>2)) ;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3499 DST(4, a0 - a1 ) ;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3500 DST(5, a6 - (a5>>2)) ;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3501 DST(6, (a2>>1) - a3 ) ;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3502 DST(7, (a4>>2) - a7 ) ;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3503 }
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3504
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3505 static int dct264_sad8x8_c(/*MpegEncContext*/ void *c, uint8_t *src1, uint8_t *src2, int stride, int h){
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3506 MpegEncContext * const s= (MpegEncContext *)c;
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3507 int16_t dct[8][8];
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3508 int i;
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3509 int sum=0;
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3510
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3511 s->dsp.diff_pixels(dct, src1, src2, stride);
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3512
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3513 #define SRC(x) dct[i][x]
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3514 #define DST(x,v) dct[i][x]= v
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3515 for( i = 0; i < 8; i++ )
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3516 DCT8_1D
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3517 #undef SRC
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3518 #undef DST
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3519
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3520 #define SRC(x) dct[x][i]
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3521 #define DST(x,v) sum += ABS(v)
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3522 for( i = 0; i < 8; i++ )
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3523 DCT8_1D
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3524 #undef SRC
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3525 #undef DST
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3526 return sum;
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3527 }
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3528 #endif
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3529
2382
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3530 static int dct_max8x8_c(/*MpegEncContext*/ void *c, uint8_t *src1, uint8_t *src2, int stride, int h){
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3531 MpegEncContext * const s= (MpegEncContext *)c;
3089
072dbc669253 MSVC-compatible __align8/__align16 declaration
diego
parents: 3036
diff changeset
3532 DECLARE_ALIGNED_8(uint64_t, aligned_temp[sizeof(DCTELEM)*64/8]);
2382
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3533 DCTELEM * const temp= (DCTELEM*)aligned_temp;
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3534 int sum=0, i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3535
2382
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3536 assert(h==8);
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3537
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3538 s->dsp.diff_pixels(temp, src1, src2, stride);
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3539 s->dsp.fdct(temp);
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3540
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3541 for(i=0; i<64; i++)
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3542 sum= FFMAX(sum, ABS(temp[i]));
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3543
2382
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3544 return sum;
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3545 }
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3546
1008
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 1007
diff changeset
3547 void simple_idct(DCTELEM *block); //FIXME
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3548
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3549 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
3550 MpegEncContext * const s= (MpegEncContext *)c;
3089
072dbc669253 MSVC-compatible __align8/__align16 declaration
diego
parents: 3036
diff changeset
3551 DECLARE_ALIGNED_8 (uint64_t, aligned_temp[sizeof(DCTELEM)*64*2/8]);
1016
michaelni
parents: 1013
diff changeset
3552 DCTELEM * const temp= (DCTELEM*)aligned_temp;
michaelni
parents: 1013
diff changeset
3553 DCTELEM * const bak = ((DCTELEM*)aligned_temp)+64;
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3554 int sum=0, i;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3555
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3556 assert(h==8);
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3557 s->mb_intra=0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3558
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3559 s->dsp.diff_pixels(temp, src1, src2, stride);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3560
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3561 memcpy(bak, temp, 64*sizeof(DCTELEM));
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3562
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3563 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
3564 s->dct_unquantize_inter(s, temp, 0, s->qscale);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3565 simple_idct(temp); //FIXME
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3566
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3567 for(i=0; i<64; i++)
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3568 sum+= (temp[i]-bak[i])*(temp[i]-bak[i]);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3569
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3570 return sum;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3571 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3572
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3573 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
3574 MpegEncContext * const s= (MpegEncContext *)c;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
3575 const uint8_t *scantable= s->intra_scantable.permutated;
3089
072dbc669253 MSVC-compatible __align8/__align16 declaration
diego
parents: 3036
diff changeset
3576 DECLARE_ALIGNED_8 (uint64_t, aligned_temp[sizeof(DCTELEM)*64/8]);
072dbc669253 MSVC-compatible __align8/__align16 declaration
diego
parents: 3036
diff changeset
3577 DECLARE_ALIGNED_8 (uint64_t, aligned_bak[stride]);
1016
michaelni
parents: 1013
diff changeset
3578 DCTELEM * const temp= (DCTELEM*)aligned_temp;
michaelni
parents: 1013
diff changeset
3579 uint8_t * const bak= (uint8_t*)aligned_bak;
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3580 int i, last, run, bits, level, distoration, start_i;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3581 const int esc_length= s->ac_esc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3582 uint8_t * length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3583 uint8_t * last_length;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3584
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3585 assert(h==8);
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3586
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3587 for(i=0; i<8; i++){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3588 ((uint32_t*)(bak + i*stride))[0]= ((uint32_t*)(src2 + i*stride))[0];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3589 ((uint32_t*)(bak + i*stride))[1]= ((uint32_t*)(src2 + i*stride))[1];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3590 }
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3591
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3592 s->dsp.diff_pixels(temp, src1, src2, stride);
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3593
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3594 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
3595
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3596 bits=0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3597
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3598 if (s->mb_intra) {
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3599 start_i = 1;
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3600 length = s->intra_ac_vlc_length;
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3601 last_length= s->intra_ac_vlc_last_length;
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3602 bits+= s->luma_dc_vlc_length[temp[0] + 256]; //FIXME chroma
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3603 } else {
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3604 start_i = 0;
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3605 length = s->inter_ac_vlc_length;
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3606 last_length= s->inter_ac_vlc_last_length;
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3607 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3608
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3609 if(last>=start_i){
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3610 run=0;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3611 for(i=start_i; i<last; i++){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3612 int j= scantable[i];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3613 level= temp[j];
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3614
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3615 if(level){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3616 level+=64;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3617 if((level&(~127)) == 0){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3618 bits+= length[UNI_AC_ENC_INDEX(run, level)];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3619 }else
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3620 bits+= esc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3621 run=0;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3622 }else
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3623 run++;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3624 }
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3625 i= scantable[last];
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3626
1011
michaelni
parents: 1008
diff changeset
3627 level= temp[i] + 64;
michaelni
parents: 1008
diff changeset
3628
michaelni
parents: 1008
diff changeset
3629 assert(level - 64);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3630
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3631 if((level&(~127)) == 0){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3632 bits+= last_length[UNI_AC_ENC_INDEX(run, level)];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3633 }else
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3634 bits+= esc_length;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3635
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3636 }
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3637
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3638 if(last>=0){
1689
1a2db2073848 split intra / inter dequantization
michael
parents: 1645
diff changeset
3639 if(s->mb_intra)
1a2db2073848 split intra / inter dequantization
michael
parents: 1645
diff changeset
3640 s->dct_unquantize_intra(s, temp, 0, s->qscale);
1a2db2073848 split intra / inter dequantization
michael
parents: 1645
diff changeset
3641 else
1a2db2073848 split intra / inter dequantization
michael
parents: 1645
diff changeset
3642 s->dct_unquantize_inter(s, temp, 0, s->qscale);
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3643 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3644
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3645 s->dsp.idct_add(bak, stride, temp);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3646
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3647 distoration= s->dsp.sse[1](NULL, bak, src1, stride, 8);
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3648
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3649 return distoration + ((bits*s->qscale*s->qscale*109 + 64)>>7);
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3650 }
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3651
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3652 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
3653 MpegEncContext * const s= (MpegEncContext *)c;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
3654 const uint8_t *scantable= s->intra_scantable.permutated;
3089
072dbc669253 MSVC-compatible __align8/__align16 declaration
diego
parents: 3036
diff changeset
3655 DECLARE_ALIGNED_8 (uint64_t, aligned_temp[sizeof(DCTELEM)*64/8]);
1016
michaelni
parents: 1013
diff changeset
3656 DCTELEM * const temp= (DCTELEM*)aligned_temp;
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3657 int i, last, run, bits, level, start_i;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3658 const int esc_length= s->ac_esc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3659 uint8_t * length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3660 uint8_t * last_length;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3661
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3662 assert(h==8);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3663
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3664 s->dsp.diff_pixels(temp, src1, src2, stride);
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3665
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3666 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
3667
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3668 bits=0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3669
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3670 if (s->mb_intra) {
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3671 start_i = 1;
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3672 length = s->intra_ac_vlc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3673 last_length= s->intra_ac_vlc_last_length;
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3674 bits+= s->luma_dc_vlc_length[temp[0] + 256]; //FIXME chroma
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3675 } else {
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3676 start_i = 0;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3677 length = s->inter_ac_vlc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3678 last_length= s->inter_ac_vlc_last_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3679 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3680
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3681 if(last>=start_i){
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3682 run=0;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3683 for(i=start_i; i<last; i++){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3684 int j= scantable[i];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3685 level= temp[j];
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3686
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3687 if(level){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3688 level+=64;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3689 if((level&(~127)) == 0){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3690 bits+= length[UNI_AC_ENC_INDEX(run, level)];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3691 }else
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3692 bits+= esc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3693 run=0;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3694 }else
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3695 run++;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3696 }
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3697 i= scantable[last];
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3698
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3699 level= temp[i] + 64;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3700
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3701 assert(level - 64);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3702
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3703 if((level&(~127)) == 0){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3704 bits+= last_length[UNI_AC_ENC_INDEX(run, level)];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3705 }else
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3706 bits+= esc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3707 }
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3708
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3709 return bits;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3710 }
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3711
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3712 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
3713 int score=0;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3714 int x,y;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3715
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3716 for(y=1; y<h; y++){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3717 for(x=0; x<16; x+=4){
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3718 score+= ABS(s[x ] - s[x +stride]) + ABS(s[x+1] - s[x+1+stride])
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3719 +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
3720 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3721 s+= stride;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3722 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3723
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3724 return score;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3725 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3726
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3727 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
3728 int score=0;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3729 int x,y;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3730
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3731 for(y=1; y<h; y++){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3732 for(x=0; x<16; x++){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3733 score+= ABS(s1[x ] - s2[x ] - s1[x +stride] + s2[x +stride]);
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3734 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3735 s1+= stride;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3736 s2+= stride;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3737 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3738
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3739 return score;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3740 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3741
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3742 #define SQ(a) ((a)*(a))
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3743 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
3744 int score=0;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3745 int x,y;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3746
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3747 for(y=1; y<h; y++){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3748 for(x=0; x<16; x+=4){
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3749 score+= SQ(s[x ] - s[x +stride]) + SQ(s[x+1] - s[x+1+stride])
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3750 +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
3751 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3752 s+= stride;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3753 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3754
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3755 return score;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3756 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3757
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3758 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
3759 int score=0;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3760 int x,y;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3761
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3762 for(y=1; y<h; y++){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3763 for(x=0; x<16; x++){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3764 score+= SQ(s1[x ] - s2[x ] - s1[x +stride] + s2[x +stride]);
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3765 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3766 s1+= stride;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3767 s2+= stride;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3768 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3769
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3770 return score;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3771 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3772
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3773 WARPER8_16_SQ(hadamard8_diff8x8_c, hadamard8_diff16_c)
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3774 WARPER8_16_SQ(hadamard8_intra8x8_c, hadamard8_intra16_c)
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3775 WARPER8_16_SQ(dct_sad8x8_c, dct_sad16_c)
3013
b6a90fabaa76 ugly, but necessary
melanson
parents: 3010
diff changeset
3776 #ifdef CONFIG_GPL
3010
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3777 WARPER8_16_SQ(dct264_sad8x8_c, dct264_sad16_c)
3013
b6a90fabaa76 ugly, but necessary
melanson
parents: 3010
diff changeset
3778 #endif
2382
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3779 WARPER8_16_SQ(dct_max8x8_c, dct_max16_c)
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3780 WARPER8_16_SQ(quant_psnr8x8_c, quant_psnr16_c)
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3781 WARPER8_16_SQ(rd8x8_c, rd16_c)
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3782 WARPER8_16_SQ(bit8x8_c, bit16_c)
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3783
3568
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3784 static void vector_fmul_c(float *dst, const float *src, int len){
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3785 int i;
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3786 for(i=0; i<len; i++)
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3787 dst[i] *= src[i];
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3788 }
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3789
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3790 static void vector_fmul_reverse_c(float *dst, const float *src0, const float *src1, int len){
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3791 int i;
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3792 src1 += len-1;
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3793 for(i=0; i<len; i++)
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3794 dst[i] = src0[i] * src1[-i];
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3795 }
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3796
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3797 void ff_vector_fmul_add_add_c(float *dst, const float *src0, const float *src1, const float *src2, int src3, int len, int step){
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3798 int i;
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3799 for(i=0; i<len; i++)
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3800 dst[i*step] = src0[i] * src1[i] + src2[i] + src3;
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3801 }
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3802
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3803 void ff_float_to_int16_c(int16_t *dst, const float *src, int len){
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3804 int i;
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3805 for(i=0; i<len; i++) {
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3806 int_fast32_t tmp = ((int32_t*)src)[i];
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3807 if(tmp & 0xf0000){
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3808 tmp = (0x43c0ffff - tmp)>>31;
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3809 // is this faster on some gcc/cpu combinations?
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3810 // if(tmp > 0x43c0ffff) tmp = 0xFFFF;
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3811 // else tmp = 0;
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3812 }
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3813 dst[i] = tmp - 0x8000;
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3814 }
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3815 }
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
3816
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3817 /* XXX: those functions should be suppressed ASAP when all IDCTs are
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3818 converted */
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3819 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
3820 {
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3821 j_rev_dct (block);
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3822 put_pixels_clamped_c(block, dest, line_size);
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3823 }
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3824 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
3825 {
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3826 j_rev_dct (block);
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3827 add_pixels_clamped_c(block, dest, line_size);
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3828 }
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3829
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3830 static void ff_jref_idct4_put(uint8_t *dest, int line_size, DCTELEM *block)
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3831 {
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3832 j_rev_dct4 (block);
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3833 put_pixels_clamped4_c(block, dest, line_size);
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3834 }
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3835 static void ff_jref_idct4_add(uint8_t *dest, int line_size, DCTELEM *block)
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3836 {
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3837 j_rev_dct4 (block);
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3838 add_pixels_clamped4_c(block, dest, line_size);
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3839 }
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3840
2257
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3841 static void ff_jref_idct2_put(uint8_t *dest, int line_size, DCTELEM *block)
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3842 {
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3843 j_rev_dct2 (block);
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3844 put_pixels_clamped2_c(block, dest, line_size);
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3845 }
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3846 static void ff_jref_idct2_add(uint8_t *dest, int line_size, DCTELEM *block)
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3847 {
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3848 j_rev_dct2 (block);
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3849 add_pixels_clamped2_c(block, dest, line_size);
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3850 }
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3851
2259
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3852 static void ff_jref_idct1_put(uint8_t *dest, int line_size, DCTELEM *block)
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3853 {
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3854 uint8_t *cm = cropTbl + MAX_NEG_CROP;
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3855
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3856 dest[0] = cm[(block[0] + 4)>>3];
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3857 }
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3858 static void ff_jref_idct1_add(uint8_t *dest, int line_size, DCTELEM *block)
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3859 {
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3860 uint8_t *cm = cropTbl + MAX_NEG_CROP;
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3861
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3862 dest[0] = cm[dest[0] + ((block[0] + 4)>>3)];
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3863 }
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3864
3215
06f98047ff26 prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents: 3199
diff changeset
3865 static void just_return() { return; }
06f98047ff26 prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents: 3199
diff changeset
3866
1201
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3867 /* init static data */
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3868 void dsputil_static_init(void)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
3869 {
751
cbe316f082bc warning fixes
michaelni
parents: 706
diff changeset
3870 int i;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
3871
1201
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3872 for(i=0;i<256;i++) cropTbl[i + MAX_NEG_CROP] = i;
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3873 for(i=0;i<MAX_NEG_CROP;i++) {
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3874 cropTbl[i] = 0;
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3875 cropTbl[i + MAX_NEG_CROP + 256] = 255;
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3876 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3877
1201
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3878 for(i=0;i<512;i++) {
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3879 squareTbl[i] = (i - 256) * (i - 256);
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3880 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3881
1201
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3882 for(i=0; i<64; i++) inv_zigzag_direct16[ff_zigzag_direct[i]]= i+1;
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3883 }
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
3884
861
243cc33da3eb * init for inv_zigzag_direct16 moved to init block
kabi
parents: 860
diff changeset
3885
1201
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3886 void dsputil_init(DSPContext* c, AVCodecContext *avctx)
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3887 {
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
3888 int i;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
3889
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3890 #ifdef CONFIG_ENCODERS
1567
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
3891 if(avctx->dct_algo==FF_DCT_FASTINT) {
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3892 c->fdct = fdct_ifast;
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
3893 c->fdct248 = fdct_ifast248;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3894 }
1567
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
3895 else if(avctx->dct_algo==FF_DCT_FAAN) {
1557
5d53c03186a1 floating point AAN DCT
michael
parents: 1527
diff changeset
3896 c->fdct = ff_faandct;
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
3897 c->fdct248 = ff_faandct248;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3898 }
1567
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
3899 else {
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3900 c->fdct = ff_jpeg_fdct_islow; //slow/accurate/default
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
3901 c->fdct248 = ff_fdct248_islow;
1567
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
3902 }
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3903 #endif //CONFIG_ENCODERS
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3904
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3905 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
3906 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
3907 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
3908 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
3909 }else{
cd43603c46f9 move h264 idct to its own file and call via function pointer in DspContext
michael
parents: 2259
diff changeset
3910 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
3911 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
3912 }
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3913 c->idct = j_rev_dct4;
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3914 c->idct_permutation_type= FF_NO_IDCT_PERM;
2257
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3915 }else if(avctx->lowres==2){
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3916 c->idct_put= ff_jref_idct2_put;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3917 c->idct_add= ff_jref_idct2_add;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3918 c->idct = j_rev_dct2;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
3919 c->idct_permutation_type= FF_NO_IDCT_PERM;
2259
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3920 }else if(avctx->lowres==3){
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3921 c->idct_put= ff_jref_idct1_put;
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3922 c->idct_add= ff_jref_idct1_add;
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3923 c->idct = j_rev_dct1;
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
3924 c->idct_permutation_type= FF_NO_IDCT_PERM;
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3925 }else{
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3926 if(avctx->idct_algo==FF_IDCT_INT){
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3927 c->idct_put= ff_jref_idct_put;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3928 c->idct_add= ff_jref_idct_add;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3929 c->idct = j_rev_dct;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3930 c->idct_permutation_type= FF_LIBMPEG2_IDCT_PERM;
2693
02925a3903b6 porting vp3 idct over to lavc idct api
michael
parents: 2651
diff changeset
3931 }else if(avctx->idct_algo==FF_IDCT_VP3){
02925a3903b6 porting vp3 idct over to lavc idct api
michael
parents: 2651
diff changeset
3932 c->idct_put= ff_vp3_idct_put_c;
02925a3903b6 porting vp3 idct over to lavc idct api
michael
parents: 2651
diff changeset
3933 c->idct_add= ff_vp3_idct_add_c;
02925a3903b6 porting vp3 idct over to lavc idct api
michael
parents: 2651
diff changeset
3934 c->idct = ff_vp3_idct_c;
02925a3903b6 porting vp3 idct over to lavc idct api
michael
parents: 2651
diff changeset
3935 c->idct_permutation_type= FF_NO_IDCT_PERM;
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3936 }else{ //accurate/default
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3937 c->idct_put= simple_idct_put;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3938 c->idct_add= simple_idct_add;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3939 c->idct = simple_idct;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3940 c->idct_permutation_type= FF_NO_IDCT_PERM;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
3941 }
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3942 }
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3943
2272
cd43603c46f9 move h264 idct to its own file and call via function pointer in DspContext
michael
parents: 2259
diff changeset
3944 c->h264_idct_add= ff_h264_idct_add_c;
2755
975074f04b95 decode H.264 with 8x8 transform.
lorenm
parents: 2707
diff changeset
3945 c->h264_idct8_add= ff_h264_idct8_add_c;
3105
2d35fb3cb940 h264: special case dc-only idct. ~1% faster overall
lorenm
parents: 3089
diff changeset
3946 c->h264_idct_dc_add= ff_h264_idct_dc_add_c;
2d35fb3cb940 h264: special case dc-only idct. ~1% faster overall
lorenm
parents: 3089
diff changeset
3947 c->h264_idct8_dc_add= ff_h264_idct8_dc_add_c;
2272
cd43603c46f9 move h264 idct to its own file and call via function pointer in DspContext
michael
parents: 2259
diff changeset
3948
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
3949 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
3950 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
3951 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
3952 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
3953 c->add_pixels_clamped = add_pixels_clamped_c;
2763
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
3954 c->add_pixels8 = add_pixels8_c;
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
3955 c->add_pixels4 = add_pixels4_c;
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
3956 c->gmc1 = gmc1_c;
3248
7aa9f80e7954 mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents: 3245
diff changeset
3957 c->gmc = ff_gmc_c;
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
3958 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
3959 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
3960 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
3961
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3962 /* TODO [0] 16 [1] 8 */
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3963 c->pix_abs[0][0] = pix_abs16_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3964 c->pix_abs[0][1] = pix_abs16_x2_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3965 c->pix_abs[0][2] = pix_abs16_y2_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3966 c->pix_abs[0][3] = pix_abs16_xy2_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3967 c->pix_abs[1][0] = pix_abs8_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3968 c->pix_abs[1][1] = pix_abs8_x2_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3969 c->pix_abs[1][2] = pix_abs8_y2_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3970 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
3971
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3972 #define dspfunc(PFX, IDX, NUM) \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3973 c->PFX ## _pixels_tab[IDX][0] = PFX ## _pixels ## NUM ## _c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3974 c->PFX ## _pixels_tab[IDX][1] = PFX ## _pixels ## NUM ## _x2_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3975 c->PFX ## _pixels_tab[IDX][2] = PFX ## _pixels ## NUM ## _y2_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3976 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
3977
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3978 dspfunc(put, 0, 16);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3979 dspfunc(put_no_rnd, 0, 16);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3980 dspfunc(put, 1, 8);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3981 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
3982 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
3983 dspfunc(put, 3, 2);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
3984
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3985 dspfunc(avg, 0, 16);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3986 dspfunc(avg_no_rnd, 0, 16);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3987 dspfunc(avg, 1, 8);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3988 dspfunc(avg_no_rnd, 1, 8);
1319
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
3989 dspfunc(avg, 2, 4);
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
3990 dspfunc(avg, 3, 2);
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
3991 #undef dspfunc
857
b510a7b6decd 100l for kabi (fixing qpel)
michaelni
parents: 853
diff changeset
3992
1864
9a73e6dc5d68 fix motion compensation with (x+1/2,y+1/2) MVs
michael
parents: 1784
diff changeset
3993 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
3994 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
3995
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
3996 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
3997 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
3998 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
3999 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
4000 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
4001 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
4002 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
4003 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
4004 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
4005
1319
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
4006 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
4007 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
4008 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
4009 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
4010 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
4011 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
4012 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
4013 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
4014 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
4015
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4016 #define dspfunc(PFX, IDX, NUM) \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4017 c->PFX ## _pixels_tab[IDX][ 0] = PFX ## NUM ## _mc00_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4018 c->PFX ## _pixels_tab[IDX][ 1] = PFX ## NUM ## _mc10_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4019 c->PFX ## _pixels_tab[IDX][ 2] = PFX ## NUM ## _mc20_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4020 c->PFX ## _pixels_tab[IDX][ 3] = PFX ## NUM ## _mc30_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4021 c->PFX ## _pixels_tab[IDX][ 4] = PFX ## NUM ## _mc01_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4022 c->PFX ## _pixels_tab[IDX][ 5] = PFX ## NUM ## _mc11_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4023 c->PFX ## _pixels_tab[IDX][ 6] = PFX ## NUM ## _mc21_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4024 c->PFX ## _pixels_tab[IDX][ 7] = PFX ## NUM ## _mc31_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4025 c->PFX ## _pixels_tab[IDX][ 8] = PFX ## NUM ## _mc02_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4026 c->PFX ## _pixels_tab[IDX][ 9] = PFX ## NUM ## _mc12_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4027 c->PFX ## _pixels_tab[IDX][10] = PFX ## NUM ## _mc22_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4028 c->PFX ## _pixels_tab[IDX][11] = PFX ## NUM ## _mc32_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4029 c->PFX ## _pixels_tab[IDX][12] = PFX ## NUM ## _mc03_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4030 c->PFX ## _pixels_tab[IDX][13] = PFX ## NUM ## _mc13_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4031 c->PFX ## _pixels_tab[IDX][14] = PFX ## NUM ## _mc23_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4032 c->PFX ## _pixels_tab[IDX][15] = PFX ## NUM ## _mc33_c
857
b510a7b6decd 100l for kabi (fixing qpel)
michaelni
parents: 853
diff changeset
4033
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4034 dspfunc(put_qpel, 0, 16);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4035 dspfunc(put_no_rnd_qpel, 0, 16);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4036
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4037 dspfunc(avg_qpel, 0, 16);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4038 /* dspfunc(avg_no_rnd_qpel, 0, 16); */
857
b510a7b6decd 100l for kabi (fixing qpel)
michaelni
parents: 853
diff changeset
4039
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4040 dspfunc(put_qpel, 1, 8);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4041 dspfunc(put_no_rnd_qpel, 1, 8);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4042
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4043 dspfunc(avg_qpel, 1, 8);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4044 /* dspfunc(avg_no_rnd_qpel, 1, 8); */
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4045
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4046 dspfunc(put_h264_qpel, 0, 16);
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4047 dspfunc(put_h264_qpel, 1, 8);
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4048 dspfunc(put_h264_qpel, 2, 4);
3020
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
4049 dspfunc(put_h264_qpel, 3, 2);
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4050 dspfunc(avg_h264_qpel, 0, 16);
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4051 dspfunc(avg_h264_qpel, 1, 8);
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4052 dspfunc(avg_h264_qpel, 2, 4);
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4053
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4054 #undef dspfunc
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4055 c->put_h264_chroma_pixels_tab[0]= put_h264_chroma_mc8_c;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4056 c->put_h264_chroma_pixels_tab[1]= put_h264_chroma_mc4_c;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4057 c->put_h264_chroma_pixels_tab[2]= put_h264_chroma_mc2_c;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4058 c->avg_h264_chroma_pixels_tab[0]= avg_h264_chroma_mc8_c;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4059 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_c;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4060 c->avg_h264_chroma_pixels_tab[2]= avg_h264_chroma_mc2_c;
3663
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
4061 c->put_no_rnd_h264_chroma_pixels_tab[0]= put_no_rnd_h264_chroma_mc8_c;
857
b510a7b6decd 100l for kabi (fixing qpel)
michaelni
parents: 853
diff changeset
4062
2415
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4063 c->weight_h264_pixels_tab[0]= weight_h264_pixels16x16_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4064 c->weight_h264_pixels_tab[1]= weight_h264_pixels16x8_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4065 c->weight_h264_pixels_tab[2]= weight_h264_pixels8x16_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4066 c->weight_h264_pixels_tab[3]= weight_h264_pixels8x8_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4067 c->weight_h264_pixels_tab[4]= weight_h264_pixels8x4_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4068 c->weight_h264_pixels_tab[5]= weight_h264_pixels4x8_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4069 c->weight_h264_pixels_tab[6]= weight_h264_pixels4x4_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4070 c->weight_h264_pixels_tab[7]= weight_h264_pixels4x2_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4071 c->weight_h264_pixels_tab[8]= weight_h264_pixels2x4_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4072 c->weight_h264_pixels_tab[9]= weight_h264_pixels2x2_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4073 c->biweight_h264_pixels_tab[0]= biweight_h264_pixels16x16_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4074 c->biweight_h264_pixels_tab[1]= biweight_h264_pixels16x8_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4075 c->biweight_h264_pixels_tab[2]= biweight_h264_pixels8x16_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4076 c->biweight_h264_pixels_tab[3]= biweight_h264_pixels8x8_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4077 c->biweight_h264_pixels_tab[4]= biweight_h264_pixels8x4_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4078 c->biweight_h264_pixels_tab[5]= biweight_h264_pixels4x8_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4079 c->biweight_h264_pixels_tab[6]= biweight_h264_pixels4x4_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4080 c->biweight_h264_pixels_tab[7]= biweight_h264_pixels4x2_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4081 c->biweight_h264_pixels_tab[8]= biweight_h264_pixels2x4_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4082 c->biweight_h264_pixels_tab[9]= biweight_h264_pixels2x2_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4083
3432
ab49baf4adad Fix compilation with --disable-decoders.
diego
parents: 3395
diff changeset
4084 #ifdef CONFIG_CAVS_DECODER
3395
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
4085 ff_cavsdsp_init(c,avctx);
3432
ab49baf4adad Fix compilation with --disable-decoders.
diego
parents: 3395
diff changeset
4086 #endif
3526
7dc8e4a12105 New functions in DSPContext for VC-1 decoding
kostya
parents: 3432
diff changeset
4087 #if defined(CONFIG_VC1_DECODER) || defined(CONFIG_WMV3_DECODER)
7dc8e4a12105 New functions in DSPContext for VC-1 decoding
kostya
parents: 3432
diff changeset
4088 ff_vc1dsp_init(c,avctx);
7dc8e4a12105 New functions in DSPContext for VC-1 decoding
kostya
parents: 3432
diff changeset
4089 #endif
3395
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
4090
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
4091 c->put_mspel_pixels_tab[0]= put_mspel8_mc00_c;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
4092 c->put_mspel_pixels_tab[1]= put_mspel8_mc10_c;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
4093 c->put_mspel_pixels_tab[2]= put_mspel8_mc20_c;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
4094 c->put_mspel_pixels_tab[3]= put_mspel8_mc30_c;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
4095 c->put_mspel_pixels_tab[4]= put_mspel8_mc02_c;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
4096 c->put_mspel_pixels_tab[5]= put_mspel8_mc12_c;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
4097 c->put_mspel_pixels_tab[6]= put_mspel8_mc22_c;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
4098 c->put_mspel_pixels_tab[7]= put_mspel8_mc32_c;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4099
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4100 #define SET_CMP_FUNC(name) \
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4101 c->name[0]= name ## 16_c;\
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4102 c->name[1]= name ## 8x8_c;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4103
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4104 SET_CMP_FUNC(hadamard8_diff)
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4105 c->hadamard8_diff[4]= hadamard8_intra16_c;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4106 SET_CMP_FUNC(dct_sad)
2382
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
4107 SET_CMP_FUNC(dct_max)
3013
b6a90fabaa76 ugly, but necessary
melanson
parents: 3010
diff changeset
4108 #ifdef CONFIG_GPL
3010
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
4109 SET_CMP_FUNC(dct264_sad)
3013
b6a90fabaa76 ugly, but necessary
melanson
parents: 3010
diff changeset
4110 #endif
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4111 c->sad[0]= pix_abs16_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4112 c->sad[1]= pix_abs8_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4113 c->sse[0]= sse16_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4114 c->sse[1]= sse8_c;
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
4115 c->sse[2]= sse4_c;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4116 SET_CMP_FUNC(quant_psnr)
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4117 SET_CMP_FUNC(rd)
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4118 SET_CMP_FUNC(bit)
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4119 c->vsad[0]= vsad16_c;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4120 c->vsad[4]= vsad_intra16_c;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4121 c->vsse[0]= vsse16_c;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4122 c->vsse[4]= vsse_intra16_c;
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
4123 c->nsse[0]= nsse16_c;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
4124 c->nsse[1]= nsse8_c;
3373
b8996cc5ccae Disable w53 and w97 cmp methods when snow encoder is disabled
gpoirier
parents: 3323
diff changeset
4125 #ifdef CONFIG_SNOW_ENCODER
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
4126 c->w53[0]= w53_16_c;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
4127 c->w53[1]= w53_8_c;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
4128 c->w97[0]= w97_16_c;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
4129 c->w97[1]= w97_8_c;
3373
b8996cc5ccae Disable w53 and w97 cmp methods when snow encoder is disabled
gpoirier
parents: 3323
diff changeset
4130 #endif
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
4131
866
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
4132 c->add_bytes= add_bytes_c;
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
4133 c->diff_bytes= diff_bytes_c;
1527
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
4134 c->sub_hfyu_median_prediction= sub_hfyu_median_prediction_c;
1273
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
4135 c->bswap_buf= bswap_buf;
2633
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
4136
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
4137 c->h264_v_loop_filter_luma= h264_v_loop_filter_luma_c;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
4138 c->h264_h_loop_filter_luma= h264_h_loop_filter_luma_c;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
4139 c->h264_v_loop_filter_chroma= h264_v_loop_filter_chroma_c;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
4140 c->h264_h_loop_filter_chroma= h264_h_loop_filter_chroma_c;
2707
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
4141 c->h264_v_loop_filter_chroma_intra= h264_v_loop_filter_chroma_intra_c;
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
4142 c->h264_h_loop_filter_chroma_intra= h264_h_loop_filter_chroma_intra_c;
3645
47821be55b6c mmx implementation of deblocking strength decision.
lorenm
parents: 3568
diff changeset
4143 c->h264_loop_filter_strength= NULL;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4144
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
4145 c->h263_h_loop_filter= h263_h_loop_filter_c;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
4146 c->h263_v_loop_filter= h263_v_loop_filter_c;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4147
2045
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
4148 c->h261_loop_filter= h261_loop_filter_c;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4149
1784
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
4150 c->try_8x8basis= try_8x8basis_c;
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
4151 c->add_8x8basis= add_8x8basis_c;
866
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
4152
3199
1651e69b9f7a 10l: Only set *compose97i *add_yblock to dsputils context if we are building with Snow enabled
gpoirier
parents: 3198
diff changeset
4153 #ifdef CONFIG_SNOW_ENCODER
3198
6b9f0c4fbdbe First part of a series of speed-enchancing patches.
gpoirier
parents: 3105
diff changeset
4154 c->vertical_compose97i = ff_snow_vertical_compose97i;
6b9f0c4fbdbe First part of a series of speed-enchancing patches.
gpoirier
parents: 3105
diff changeset
4155 c->horizontal_compose97i = ff_snow_horizontal_compose97i;
6b9f0c4fbdbe First part of a series of speed-enchancing patches.
gpoirier
parents: 3105
diff changeset
4156 c->inner_add_yblock = ff_snow_inner_add_yblock;
3199
1651e69b9f7a 10l: Only set *compose97i *add_yblock to dsputils context if we are building with Snow enabled
gpoirier
parents: 3198
diff changeset
4157 #endif
3198
6b9f0c4fbdbe First part of a series of speed-enchancing patches.
gpoirier
parents: 3105
diff changeset
4158
3536
545a15c19c91 sse & sse2 implementations of vorbis channel coupling.
lorenm
parents: 3526
diff changeset
4159 #ifdef CONFIG_VORBIS_DECODER
545a15c19c91 sse & sse2 implementations of vorbis channel coupling.
lorenm
parents: 3526
diff changeset
4160 c->vorbis_inverse_coupling = vorbis_inverse_coupling;
545a15c19c91 sse & sse2 implementations of vorbis channel coupling.
lorenm
parents: 3526
diff changeset
4161 #endif
3568
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4162 c->vector_fmul = vector_fmul_c;
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4163 c->vector_fmul_reverse = vector_fmul_reverse_c;
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4164 c->vector_fmul_add_add = ff_vector_fmul_add_add_c;
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4165 c->float_to_int16 = ff_float_to_int16_c;
3536
545a15c19c91 sse & sse2 implementations of vorbis channel coupling.
lorenm
parents: 3526
diff changeset
4166
3245
c2c29be6282e remove mpegvideo.c img resample dependancy
michael
parents: 3215
diff changeset
4167 c->shrink[0]= ff_img_copy_plane;
c2c29be6282e remove mpegvideo.c img resample dependancy
michael
parents: 3215
diff changeset
4168 c->shrink[1]= ff_shrink22;
c2c29be6282e remove mpegvideo.c img resample dependancy
michael
parents: 3215
diff changeset
4169 c->shrink[2]= ff_shrink44;
c2c29be6282e remove mpegvideo.c img resample dependancy
michael
parents: 3215
diff changeset
4170 c->shrink[3]= ff_shrink88;
c2c29be6282e remove mpegvideo.c img resample dependancy
michael
parents: 3215
diff changeset
4171
3215
06f98047ff26 prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents: 3199
diff changeset
4172 c->prefetch= just_return;
06f98047ff26 prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents: 3199
diff changeset
4173
2
2e2c46c87460 fixed config for direct mplayer build compatibility
glantau
parents: 0
diff changeset
4174 #ifdef HAVE_MMX
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4175 dsputil_init_mmx(c, avctx);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
4176 #endif
62
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
4177 #ifdef ARCH_ARMV4L
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4178 dsputil_init_armv4l(c, avctx);
62
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
4179 #endif
88
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
4180 #ifdef HAVE_MLIB
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4181 dsputil_init_mlib(c, avctx);
88
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
4182 #endif
1959
55b7435c59b8 VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents: 1866
diff changeset
4183 #ifdef ARCH_SPARC
55b7435c59b8 VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents: 1866
diff changeset
4184 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
4185 #endif
214
73df666cacc7 Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents: 209
diff changeset
4186 #ifdef ARCH_ALPHA
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4187 dsputil_init_alpha(c, avctx);
214
73df666cacc7 Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents: 209
diff changeset
4188 #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
4189 #ifdef ARCH_POWERPC
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4190 dsputil_init_ppc(c, avctx);
626
23a093d6e450 patch by Heliodoro Tammaro <helio at interactives dot org>
michaelni
parents: 625
diff changeset
4191 #endif
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 676
diff changeset
4192 #ifdef HAVE_MMI
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4193 dsputil_init_mmi(c, avctx);
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 676
diff changeset
4194 #endif
1259
e8c3884f2c7e sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents: 1201
diff changeset
4195 #ifdef ARCH_SH4
e8c3884f2c7e sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents: 1201
diff changeset
4196 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
4197 #endif
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4198
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4199 switch(c->idct_permutation_type){
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4200 case FF_NO_IDCT_PERM:
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4201 for(i=0; i<64; i++)
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4202 c->idct_permutation[i]= i;
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4203 break;
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4204 case FF_LIBMPEG2_IDCT_PERM:
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4205 for(i=0; i<64; i++)
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4206 c->idct_permutation[i]= (i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2);
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4207 break;
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4208 case FF_SIMPLE_IDCT_PERM:
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4209 for(i=0; i<64; i++)
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4210 c->idct_permutation[i]= simple_mmx_permutation[i];
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4211 break;
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4212 case FF_TRANSPOSE_IDCT_PERM:
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4213 for(i=0; i<64; i++)
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4214 c->idct_permutation[i]= ((i&7)<<3) | (i>>3);
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4215 break;
2696
9699d325049d porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents: 2693
diff changeset
4216 case FF_PARTTRANS_IDCT_PERM:
9699d325049d porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents: 2693
diff changeset
4217 for(i=0; i<64; i++)
9699d325049d porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents: 2693
diff changeset
4218 c->idct_permutation[i]= (i&0x24) | ((i&3)<<3) | ((i>>3)&3);
9699d325049d porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents: 2693
diff changeset
4219 break;
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4220 default:
1598
932d306bf1dc av_log() patch by (Michel Bardiaux <mbardiaux at peaktime dot be>)
michael
parents: 1571
diff changeset
4221 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
4222 }
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
4223 }
252
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
4224