annotate dsputil.c @ 11376:0b220468ba0d libavcodec

Move prototypes for various dsputil init functions to dsputil.h
author mru
date Sat, 06 Mar 2010 22:36:49 +0000
parents 84963c795459
children f46b68960464
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
8629
04423b2f6e0b cosmetics: Remove pointless period after copyright statement non-sentences.
diego
parents: 8627
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 *
5214
470601203f44 Group all copyright and author notices together.
diego
parents: 5151
diff changeset
6 * gmc & q-pel & 32/64 bit based MC by Michael Niedermayer <michaelni@gmx.at>
470601203f44 Group all copyright and author notices together.
diego
parents: 5151
diff changeset
7 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3807
diff changeset
8 * This file is part of FFmpeg.
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3807
diff changeset
9 *
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3807
diff changeset
10 * FFmpeg is free software; you can redistribute it and/or
429
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
11 * modify it under the terms of the GNU Lesser General Public
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
12 * License as published by the Free Software Foundation; either
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3807
diff changeset
13 * version 2.1 of the License, or (at your option) any later version.
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
14 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3807
diff changeset
15 * FFmpeg is distributed in the hope that it will be useful,
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
429
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
18 * Lesser General Public License for more details.
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
19 *
429
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
20 * You should have received a copy of the GNU Lesser General Public
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3807
diff changeset
21 * License along with FFmpeg; if not, write to the Free Software
3036
0b546eab515d Update licensing information: The FSF changed postal address.
diego
parents: 3029
diff changeset
22 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
23 */
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
24
1106
1e39f273ecd6 per file doxy
michaelni
parents: 1101
diff changeset
25 /**
8718
e9d9d946f213 Use full internal pathname in doxygen @file directives.
diego
parents: 8629
diff changeset
26 * @file libavcodec/dsputil.c
1106
1e39f273ecd6 per file doxy
michaelni
parents: 1101
diff changeset
27 * DSP utils
1e39f273ecd6 per file doxy
michaelni
parents: 1101
diff changeset
28 */
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
29
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
30 #include "avcodec.h"
986e461dc072 Initial revision
glantau
parents:
diff changeset
31 #include "dsputil.h"
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
32 #include "simple_idct.h"
1557
5d53c03186a1 floating point AAN DCT
michael
parents: 1527
diff changeset
33 #include "faandct.h"
6407
7f9e2b5893fc floating point AAN IDCT
michael
parents: 6387
diff changeset
34 #include "faanidct.h"
8627
d6bab465b82c moves mid_pred() into mathops.h (with arch specific code split by directory)
aurel
parents: 8596
diff changeset
35 #include "mathops.h"
3198
6b9f0c4fbdbe First part of a series of speed-enchancing patches.
gpoirier
parents: 3105
diff changeset
36 #include "snow.h"
10748
36611425fedb Add required header #includes for mpegvideo.h and config.h.
diego
parents: 10644
diff changeset
37 #include "mpegvideo.h"
36611425fedb Add required header #includes for mpegvideo.h and config.h.
diego
parents: 10644
diff changeset
38 #include "config.h"
11375
84963c795459 Move some prototypes from dsputil.c to reasonable header files
mru
parents: 11369
diff changeset
39 #include "lpc.h"
84963c795459 Move some prototypes from dsputil.c to reasonable header files
mru
parents: 11369
diff changeset
40 #include "ac3dec.h"
84963c795459 Move some prototypes from dsputil.c to reasonable header files
mru
parents: 11369
diff changeset
41 #include "vorbis.h"
84963c795459 Move some prototypes from dsputil.c to reasonable header files
mru
parents: 11369
diff changeset
42 #include "png.h"
676
c3bdb00a98a9 dont store version for bit-exact tests
michaelni
parents: 651
diff changeset
43
2522
e25782262d7d kill warnings patch by (Mns Rullgrd <mru inprovide com>)
michael
parents: 2448
diff changeset
44 /* snow.c */
e25782262d7d kill warnings patch by (Mns Rullgrd <mru inprovide com>)
michael
parents: 2448
diff changeset
45 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
46
8120
3d4e01bcd2a5 Electronic Arts TGQ/TQI/MAD IDCT algorithm
pross
parents: 7995
diff changeset
47 /* eaidct.c */
3d4e01bcd2a5 Electronic Arts TGQ/TQI/MAD IDCT algorithm
pross
parents: 7995
diff changeset
48 void ff_ea_idct_put_c(uint8_t *dest, int linesize, DCTELEM *block);
3d4e01bcd2a5 Electronic Arts TGQ/TQI/MAD IDCT algorithm
pross
parents: 7995
diff changeset
49
11231
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
50 /* binkidct.c */
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
51 void ff_bink_idct_c (DCTELEM *block);
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
52 void ff_bink_idct_add_c(uint8_t *dest, int linesize, DCTELEM *block);
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
53 void ff_bink_idct_put_c(uint8_t *dest, int linesize, DCTELEM *block);
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
54
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
55 uint8_t ff_cropTbl[256 + 2 * MAX_NEG_CROP] = {0, };
4179
46d38e3d7038 rename squareTbl -> ff_squareTbl
mru
parents: 4176
diff changeset
56 uint32_t ff_squareTbl[512] = {0, };
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
57
6387
922774d1fb9c cleaner way of writing packed-byte constants.
lorenm
parents: 6386
diff changeset
58 // 0x7f7f7f7f or 0x7f7f7f7f7f7f7f7f or whatever, depending on the cpu's native arithmetic size
922774d1fb9c cleaner way of writing packed-byte constants.
lorenm
parents: 6386
diff changeset
59 #define pb_7f (~0UL/255 * 0x7f)
922774d1fb9c cleaner way of writing packed-byte constants.
lorenm
parents: 6386
diff changeset
60 #define pb_80 (~0UL/255 * 0x80)
6385
40fbc878ce3f pseudo-simd add_bytes and diff_bytes
lorenm
parents: 6384
diff changeset
61
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
62 const uint8_t ff_zigzag_direct[64] = {
706
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 689
diff changeset
63 0, 1, 8, 16, 9, 2, 3, 10,
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 689
diff changeset
64 17, 24, 32, 25, 18, 11, 4, 5,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
65 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
66 27, 20, 13, 6, 7, 14, 21, 28,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
67 35, 42, 49, 56, 57, 50, 43, 36,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
68 29, 22, 15, 23, 30, 37, 44, 51,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
69 58, 59, 52, 45, 38, 31, 39, 46,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
70 53, 60, 61, 54, 47, 55, 62, 63
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
71 };
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
72
1567
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
73 /* Specific zigzag scan for 248 idct. NOTE that unlike the
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
74 specification, we interleave the fields */
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
75 const uint8_t ff_zigzag248_direct[64] = {
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
76 0, 8, 1, 9, 16, 24, 2, 10,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
77 17, 25, 32, 40, 48, 56, 33, 41,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
78 18, 26, 3, 11, 4, 12, 19, 27,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
79 34, 42, 49, 57, 50, 58, 35, 43,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
80 20, 28, 5, 13, 6, 14, 21, 29,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
81 36, 44, 51, 59, 52, 60, 37, 45,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
82 22, 30, 7, 15, 23, 31, 38, 46,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
83 53, 61, 54, 62, 39, 47, 55, 63,
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
84 };
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
85
220
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
86 /* not permutated inverse zigzag_direct + 1 for MMX quantizer */
11369
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11231
diff changeset
87 DECLARE_ALIGNED(16, uint16_t, inv_zigzag_direct16)[64];
220
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
88
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
89 const uint8_t ff_alternate_horizontal_scan[64] = {
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
90 0, 1, 2, 3, 8, 9, 16, 17,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
91 10, 11, 4, 5, 6, 7, 15, 14,
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
92 13, 12, 19, 18, 24, 25, 32, 33,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
93 26, 27, 20, 21, 22, 23, 28, 29,
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
94 30, 31, 34, 35, 40, 41, 48, 49,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
95 42, 43, 36, 37, 38, 39, 44, 45,
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
96 46, 47, 50, 51, 56, 57, 58, 59,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
97 52, 53, 54, 55, 60, 61, 62, 63,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
98 };
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
99
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
100 const uint8_t ff_alternate_vertical_scan[64] = {
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
101 0, 8, 16, 24, 1, 9, 2, 10,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
102 17, 25, 32, 40, 48, 56, 57, 49,
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
103 41, 33, 26, 18, 3, 11, 4, 12,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
104 19, 27, 34, 42, 50, 58, 35, 43,
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
105 51, 59, 20, 28, 5, 13, 6, 14,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
106 21, 29, 36, 44, 52, 60, 37, 45,
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
107 53, 61, 22, 30, 7, 15, 23, 31,
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
108 38, 46, 54, 62, 39, 47, 55, 63,
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
109 };
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
110
10207
658b2ca35e22 extend ff_inverse[], and fix its documentation
lorenm
parents: 10105
diff changeset
111 /* a*inverse[b]>>32 == a/b for all 0<=a<=16909558 && 2<=b<=256
658b2ca35e22 extend ff_inverse[], and fix its documentation
lorenm
parents: 10105
diff changeset
112 * for a>16909558, is an overestimate by less than 1 part in 1<<24 */
658b2ca35e22 extend ff_inverse[], and fix its documentation
lorenm
parents: 10105
diff changeset
113 const uint32_t ff_inverse[257]={
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
114 0, 4294967295U,2147483648U,1431655766, 1073741824, 858993460, 715827883, 613566757,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
115 536870912, 477218589, 429496730, 390451573, 357913942, 330382100, 306783379, 286331154,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
116 268435456, 252645136, 238609295, 226050911, 214748365, 204522253, 195225787, 186737709,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
117 178956971, 171798692, 165191050, 159072863, 153391690, 148102321, 143165577, 138547333,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
118 134217728, 130150525, 126322568, 122713352, 119304648, 116080198, 113025456, 110127367,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
119 107374183, 104755300, 102261127, 99882961, 97612894, 95443718, 93368855, 91382283,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
120 89478486, 87652394, 85899346, 84215046, 82595525, 81037119, 79536432, 78090315,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
121 76695845, 75350304, 74051161, 72796056, 71582789, 70409300, 69273667, 68174085,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
122 67108864, 66076420, 65075263, 64103990, 63161284, 62245903, 61356676, 60492498,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
123 59652324, 58835169, 58040099, 57266231, 56512728, 55778797, 55063684, 54366675,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
124 53687092, 53024288, 52377650, 51746594, 51130564, 50529028, 49941481, 49367441,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
125 48806447, 48258060, 47721859, 47197443, 46684428, 46182445, 45691142, 45210183,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
126 44739243, 44278014, 43826197, 43383509, 42949673, 42524429, 42107523, 41698712,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
127 41297763, 40904451, 40518560, 40139882, 39768216, 39403370, 39045158, 38693400,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
128 38347923, 38008561, 37675152, 37347542, 37025581, 36709123, 36398028, 36092163,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
129 35791395, 35495598, 35204650, 34918434, 34636834, 34359739, 34087043, 33818641,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
130 33554432, 33294321, 33038210, 32786010, 32537632, 32292988, 32051995, 31814573,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
131 31580642, 31350127, 31122952, 30899046, 30678338, 30460761, 30246249, 30034737,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
132 29826162, 29620465, 29417585, 29217465, 29020050, 28825284, 28633116, 28443493,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
133 28256364, 28071682, 27889399, 27709467, 27531842, 27356480, 27183338, 27012373,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
134 26843546, 26676816, 26512144, 26349493, 26188825, 26030105, 25873297, 25718368,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
135 25565282, 25414008, 25264514, 25116768, 24970741, 24826401, 24683721, 24542671,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
136 24403224, 24265352, 24129030, 23994231, 23860930, 23729102, 23598722, 23469767,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
137 23342214, 23216040, 23091223, 22967740, 22845571, 22724695, 22605092, 22486740,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
138 22369622, 22253717, 22139007, 22025474, 21913099, 21801865, 21691755, 21582751,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
139 21474837, 21367997, 21262215, 21157475, 21053762, 20951060, 20849356, 20748635,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
140 20648882, 20550083, 20452226, 20355296, 20259280, 20164166, 20069941, 19976593,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
141 19884108, 19792477, 19701685, 19611723, 19522579, 19434242, 19346700, 19259944,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
142 19173962, 19088744, 19004281, 18920561, 18837576, 18755316, 18673771, 18592933,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
143 18512791, 18433337, 18354562, 18276457, 18199014, 18122225, 18046082, 17970575,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
144 17895698, 17821442, 17747799, 17674763, 17602325, 17530479, 17459217, 17388532,
220
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
145 17318417, 17248865, 17179870, 17111424, 17043522, 16976156, 16909321, 16843010,
10207
658b2ca35e22 extend ff_inverse[], and fix its documentation
lorenm
parents: 10105
diff changeset
146 16777216
220
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
147 };
0b234715e205 (commit by michael)
arpi_esp
parents: 214
diff changeset
148
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
149 /* Input permutation for the simple_idct_mmx */
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
150 static const uint8_t simple_mmx_permutation[64]={
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
151 0x00, 0x08, 0x04, 0x09, 0x01, 0x0C, 0x05, 0x0D,
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
152 0x10, 0x18, 0x14, 0x19, 0x11, 0x1C, 0x15, 0x1D,
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
153 0x20, 0x28, 0x24, 0x29, 0x21, 0x2C, 0x25, 0x2D,
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
154 0x12, 0x1A, 0x16, 0x1B, 0x13, 0x1E, 0x17, 0x1F,
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
155 0x02, 0x0A, 0x06, 0x0B, 0x03, 0x0E, 0x07, 0x0F,
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
156 0x30, 0x38, 0x34, 0x39, 0x31, 0x3C, 0x35, 0x3D,
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
157 0x22, 0x2A, 0x26, 0x2B, 0x23, 0x2E, 0x27, 0x2F,
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
158 0x32, 0x3A, 0x36, 0x3B, 0x33, 0x3E, 0x37, 0x3F,
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
159 };
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
160
6600
c3213c91124c Add a new IDCT permutation, used in xvid_sse2 and possibly future similar IDCTs.
astrange
parents: 6450
diff changeset
161 static const uint8_t idct_sse2_row_perm[8] = {0, 4, 1, 5, 2, 6, 3, 7};
c3213c91124c Add a new IDCT permutation, used in xvid_sse2 and possibly future similar IDCTs.
astrange
parents: 6450
diff changeset
162
6438
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
163 void ff_init_scantable(uint8_t *permutation, ScanTable *st, const uint8_t *src_scantable){
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
164 int i;
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
165 int end;
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
166
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
167 st->scantable= src_scantable;
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
168
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
169 for(i=0; i<64; i++){
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
170 int j;
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
171 j = src_scantable[i];
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
172 st->permutated[i] = permutation[j];
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8496
diff changeset
173 #if ARCH_PPC
6438
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
174 st->inverse[j] = i;
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
175 #endif
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
176 }
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
177
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
178 end=-1;
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
179 for(i=0; i<64; i++){
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
180 int j;
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
181 j = st->permutated[i];
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
182 if(j>end) end=j;
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
183 st->raster_end[i]= end;
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
184 }
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
185 }
e1dd408a7864 move ff_init_scantable() into dsputil
aurel
parents: 6437
diff changeset
186
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
187 static int pix_sum_c(uint8_t * pix, int line_size)
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
188 {
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
189 int s, i, j;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
190
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
191 s = 0;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
192 for (i = 0; i < 16; i++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
193 for (j = 0; j < 16; j += 8) {
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
194 s += pix[0];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
195 s += pix[1];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
196 s += pix[2];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
197 s += pix[3];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
198 s += pix[4];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
199 s += pix[5];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
200 s += pix[6];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
201 s += pix[7];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
202 pix += 8;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
203 }
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
204 pix += line_size - 16;
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
205 }
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
206 return s;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
207 }
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
208
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
209 static int pix_norm1_c(uint8_t * pix, int line_size)
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
210 {
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
211 int s, i, j;
4179
46d38e3d7038 rename squareTbl -> ff_squareTbl
mru
parents: 4176
diff changeset
212 uint32_t *sq = ff_squareTbl + 256;
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
213
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
214 s = 0;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
215 for (i = 0; i < 16; i++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
216 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
217 #if 0
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
218 s += sq[pix[0]];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
219 s += sq[pix[1]];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
220 s += sq[pix[2]];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
221 s += sq[pix[3]];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
222 s += sq[pix[4]];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
223 s += sq[pix[5]];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
224 s += sq[pix[6]];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
225 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
226 #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
227 #if LONG_MAX > 2147483647
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
228 register uint64_t x=*(uint64_t*)pix;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
229 s += sq[x&0xff];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
230 s += sq[(x>>8)&0xff];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
231 s += sq[(x>>16)&0xff];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
232 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
233 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
234 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
235 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
236 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
237 #else
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
238 register uint32_t x=*(uint32_t*)pix;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
239 s += sq[x&0xff];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
240 s += sq[(x>>8)&0xff];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
241 s += sq[(x>>16)&0xff];
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
242 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
243 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
244 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
245 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
246 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
247 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
248 #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
249 #endif
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
250 pix += 8;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
251 }
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
252 pix += line_size - 16;
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
253 }
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
254 return s;
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
255 }
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
256
6241
2799f65a24de const src for bswap
michael
parents: 6218
diff changeset
257 static void bswap_buf(uint32_t *dst, const uint32_t *src, int w){
1273
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
258 int i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
259
1273
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
260 for(i=0; i+8<=w; i+=8){
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
261 dst[i+0]= bswap_32(src[i+0]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
262 dst[i+1]= bswap_32(src[i+1]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
263 dst[i+2]= bswap_32(src[i+2]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
264 dst[i+3]= bswap_32(src[i+3]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
265 dst[i+4]= bswap_32(src[i+4]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
266 dst[i+5]= bswap_32(src[i+5]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
267 dst[i+6]= bswap_32(src[i+6]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
268 dst[i+7]= bswap_32(src[i+7]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
269 }
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
270 for(;i<w; i++){
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
271 dst[i+0]= bswap_32(src[i+0]);
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
272 }
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
273 }
612
c0005de2be59 new ratecontrol code
michaelni
parents: 590
diff changeset
274
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
275 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
276 {
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
277 int s, i;
4179
46d38e3d7038 rename squareTbl -> ff_squareTbl
mru
parents: 4176
diff changeset
278 uint32_t *sq = ff_squareTbl + 256;
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
279
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
280 s = 0;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
281 for (i = 0; i < h; i++) {
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
282 s += sq[pix1[0] - pix2[0]];
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
283 s += sq[pix1[1] - pix2[1]];
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
284 s += sq[pix1[2] - pix2[2]];
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
285 s += sq[pix1[3] - pix2[3]];
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
286 pix1 += line_size;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
287 pix2 += line_size;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
288 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
289 return s;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
290 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
291
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
292 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
293 {
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
294 int s, i;
4179
46d38e3d7038 rename squareTbl -> ff_squareTbl
mru
parents: 4176
diff changeset
295 uint32_t *sq = ff_squareTbl + 256;
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
296
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
297 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
298 for (i = 0; i < h; i++) {
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
299 s += sq[pix1[0] - pix2[0]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
300 s += sq[pix1[1] - pix2[1]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
301 s += sq[pix1[2] - pix2[2]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
302 s += sq[pix1[3] - pix2[3]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
303 s += sq[pix1[4] - pix2[4]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
304 s += sq[pix1[5] - pix2[5]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
305 s += sq[pix1[6] - pix2[6]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
306 s += sq[pix1[7] - pix2[7]];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
307 pix1 += line_size;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
308 pix2 += line_size;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
309 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
310 return s;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
311 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
312
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
313 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
314 {
1012
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
315 int s, i;
4179
46d38e3d7038 rename squareTbl -> ff_squareTbl
mru
parents: 4176
diff changeset
316 uint32_t *sq = ff_squareTbl + 256;
884
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 882
diff changeset
317
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 882
diff changeset
318 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
319 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
320 s += sq[pix1[ 0] - pix2[ 0]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
321 s += sq[pix1[ 1] - pix2[ 1]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
322 s += sq[pix1[ 2] - pix2[ 2]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
323 s += sq[pix1[ 3] - pix2[ 3]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
324 s += sq[pix1[ 4] - pix2[ 4]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
325 s += sq[pix1[ 5] - pix2[ 5]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
326 s += sq[pix1[ 6] - pix2[ 6]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
327 s += sq[pix1[ 7] - pix2[ 7]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
328 s += sq[pix1[ 8] - pix2[ 8]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
329 s += sq[pix1[ 9] - pix2[ 9]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
330 s += sq[pix1[10] - pix2[10]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
331 s += sq[pix1[11] - pix2[11]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
332 s += sq[pix1[12] - pix2[12]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
333 s += sq[pix1[13] - pix2[13]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
334 s += sq[pix1[14] - pix2[14]];
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
335 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
336
1012
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
337 pix1 += line_size;
7a5038ec769b sse16_c is totally fucked up (unaligned loads, LONG_MAX is undefined,
mellum
parents: 1011
diff changeset
338 pix2 += line_size;
884
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 882
diff changeset
339 }
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 882
diff changeset
340 return s;
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 882
diff changeset
341 }
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 882
diff changeset
342
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
343
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8496
diff changeset
344 #if CONFIG_SNOW_ENCODER //dwt is in snow.c
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
345 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
346 int s, i, j;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
347 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
348 int tmp[32*32];
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
349 int level, ori;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
350 static const int scale[2][2][4][4]={
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
351 {
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
352 {
3323
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
353 // 9/7 8x8 dec=3
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
354 {268, 239, 239, 213},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
355 { 0, 224, 224, 152},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
356 { 0, 135, 135, 110},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
357 },{
3323
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
358 // 9/7 16x16 or 32x32 dec=4
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
359 {344, 310, 310, 280},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
360 { 0, 320, 320, 228},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
361 { 0, 175, 175, 136},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
362 { 0, 129, 129, 102},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
363 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
364 },{
3323
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
365 {
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
366 // 5/3 8x8 dec=3
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
367 {275, 245, 245, 218},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
368 { 0, 230, 230, 156},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
369 { 0, 138, 138, 113},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
370 },{
3323
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
371 // 5/3 16x16 or 32x32 dec=4
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
372 {352, 317, 317, 286},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
373 { 0, 328, 328, 233},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
374 { 0, 180, 180, 140},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
375 { 0, 132, 132, 105},
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
376 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
377 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
378 };
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
379
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
380 for (i = 0; i < h; i++) {
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
381 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
382 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
383 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
384 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
385 tmp[32*i+j+3] = (pix1[j+3] - pix2[j+3])<<4;
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
386 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
387 pix1 += line_size;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
388 pix2 += line_size;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
389 }
2639
e2780f828440 put most codecs under ifdefs
michael
parents: 2633
diff changeset
390
3323
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
391 ff_spatial_dwt(tmp, w, h, 32, type, dec_count);
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
392
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
393 s=0;
3323
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
394 assert(w==h);
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
395 for(level=0; level<dec_count; level++){
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
396 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
397 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
398 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
399 int stride= 32<<(dec_count-level);
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
400 int sy= (ori&2) ? stride>>1 : 0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
401
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
402 for(i=0; i<size; i++){
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
403 for(j=0; j<size; j++){
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
404 int v= tmp[sx + sy + i*stride + j] * scale[type][dec_count-3][level][ori];
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
405 s += FFABS(v);
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
406 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
407 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
408 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
409 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
410 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
411 return s>>9;
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
412 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
413
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
414 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
415 return w_c(v, pix1, pix2, line_size, 8, h, 1);
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
416 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
417
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
418 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
419 return w_c(v, pix1, pix2, line_size, 8, h, 0);
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
420 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
421
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
422 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
423 return w_c(v, pix1, pix2, line_size, 16, h, 1);
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
424 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
425
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
426 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
427 return w_c(v, pix1, pix2, line_size, 16, h, 0);
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
428 }
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
429
4197
bbe0bc387a19 revert bad checkin
mru
parents: 4196
diff changeset
430 int w53_32_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h){
3323
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
431 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
432 }
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
433
4197
bbe0bc387a19 revert bad checkin
mru
parents: 4196
diff changeset
434 int w97_32_c(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h){
3323
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
435 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
436 }
3373
b8996cc5ccae Disable w53 and w97 cmp methods when snow encoder is disabled
gpoirier
parents: 3323
diff changeset
437 #endif
3323
87c54a3f8d19 Snow: fix subband weighting in wavelet cmp functions. use 32x32 cmp in iterative motion estimation.
lorenm
parents: 3248
diff changeset
438
6437
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
439 /* draw the edges of width 'w' of an image of size width, height */
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
440 //FIXME check that this is ok for mpeg4 interlaced
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
441 static void draw_edges_c(uint8_t *buf, int wrap, int width, int height, int w)
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
442 {
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
443 uint8_t *ptr, *last_line;
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
444 int i;
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
445
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
446 last_line = buf + (height - 1) * wrap;
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
447 for(i=0;i<w;i++) {
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
448 /* top and bottom */
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
449 memcpy(buf - (i + 1) * wrap, buf, width);
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
450 memcpy(last_line + (i + 1) * wrap, last_line, width);
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
451 }
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
452 /* left and right */
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
453 ptr = buf;
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
454 for(i=0;i<height;i++) {
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
455 memset(ptr - w, ptr[0], w);
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
456 memset(ptr + width, ptr[width-1], w);
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
457 ptr += wrap;
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
458 }
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
459 /* corners */
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
460 for(i=0;i<w;i++) {
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
461 memset(buf - (i + 1) * wrap - w, buf[0], w); /* top left */
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
462 memset(buf - (i + 1) * wrap + width, buf[width-1], w); /* top right */
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
463 memset(last_line + (i + 1) * wrap - w, last_line[0], w); /* top left */
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
464 memset(last_line + (i + 1) * wrap + width, last_line[width-1], w); /* top right */
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
465 }
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
466 }
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
467
6445
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
468 /**
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
469 * Copies a rectangular area of samples to a temporary buffer and replicates the boarder samples.
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
470 * @param buf destination buffer
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
471 * @param src source buffer
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
472 * @param linesize number of bytes between 2 vertically adjacent samples in both the source and destination buffers
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
473 * @param block_w width of block
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
474 * @param block_h height of block
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
475 * @param src_x x coordinate of the top left sample of the block in the source buffer
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
476 * @param src_y y coordinate of the top left sample of the block in the source buffer
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
477 * @param w width of the source buffer
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
478 * @param h height of the source buffer
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
479 */
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
480 void ff_emulated_edge_mc(uint8_t *buf, uint8_t *src, int linesize, int block_w, int block_h,
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
481 int src_x, int src_y, int w, int h){
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
482 int x, y;
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
483 int start_y, start_x, end_y, end_x;
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
484
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
485 if(src_y>= h){
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
486 src+= (h-1-src_y)*linesize;
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
487 src_y=h-1;
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
488 }else if(src_y<=-block_h){
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
489 src+= (1-block_h-src_y)*linesize;
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
490 src_y=1-block_h;
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
491 }
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
492 if(src_x>= w){
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
493 src+= (w-1-src_x);
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
494 src_x=w-1;
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
495 }else if(src_x<=-block_w){
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
496 src+= (1-block_w-src_x);
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
497 src_x=1-block_w;
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
498 }
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
499
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
500 start_y= FFMAX(0, -src_y);
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
501 start_x= FFMAX(0, -src_x);
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
502 end_y= FFMIN(block_h, h-src_y);
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
503 end_x= FFMIN(block_w, w-src_x);
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
504
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
505 // copy existing part
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
506 for(y=start_y; y<end_y; y++){
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
507 for(x=start_x; x<end_x; x++){
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
508 buf[x + y*linesize]= src[x + y*linesize];
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
509 }
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
510 }
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
511
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
512 //top
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
513 for(y=0; y<start_y; y++){
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
514 for(x=start_x; x<end_x; x++){
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
515 buf[x + y*linesize]= buf[x + start_y*linesize];
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
516 }
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
517 }
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
518
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
519 //bottom
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
520 for(y=end_y; y<block_h; y++){
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
521 for(x=start_x; x<end_x; x++){
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
522 buf[x + y*linesize]= buf[x + (end_y-1)*linesize];
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
523 }
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
524 }
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
525
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
526 for(y=0; y<block_h; y++){
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
527 //left
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
528 for(x=0; x<start_x; x++){
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
529 buf[x + y*linesize]= buf[start_x + y*linesize];
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
530 }
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
531
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
532 //right
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
533 for(x=end_x; x<block_w; x++){
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
534 buf[x + y*linesize]= buf[end_x - 1 + y*linesize];
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
535 }
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
536 }
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
537 }
2b553c57ec51 move ff_emulated_edge_mc() to dsputil
aurel
parents: 6438
diff changeset
538
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
539 static void get_pixels_c(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
540 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
541 int i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
542
986e461dc072 Initial revision
glantau
parents:
diff changeset
543 /* read the pixels */
986e461dc072 Initial revision
glantau
parents:
diff changeset
544 for(i=0;i<8;i++) {
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
545 block[0] = pixels[0];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
546 block[1] = pixels[1];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
547 block[2] = pixels[2];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
548 block[3] = pixels[3];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
549 block[4] = pixels[4];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
550 block[5] = pixels[5];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
551 block[6] = pixels[6];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
552 block[7] = pixels[7];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
553 pixels += line_size;
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
554 block += 8;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
555 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
556 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
557
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
558 static void diff_pixels_c(DCTELEM *restrict block, const uint8_t *s1,
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
559 const uint8_t *s2, int stride){
324
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
560 int i;
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
561
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
562 /* read the pixels */
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
563 for(i=0;i<8;i++) {
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
564 block[0] = s1[0] - s2[0];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
565 block[1] = s1[1] - s2[1];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
566 block[2] = s1[2] - s2[2];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
567 block[3] = s1[3] - s2[3];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
568 block[4] = s1[4] - s2[4];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
569 block[5] = s1[5] - s2[5];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
570 block[6] = s1[6] - s2[6];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
571 block[7] = s1[7] - s2[7];
324
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
572 s1 += stride;
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
573 s2 += stride;
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
574 block += 8;
324
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
575 }
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
576 }
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
577
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 320
diff changeset
578
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
579 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
580 int line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
581 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
582 int i;
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
583 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
584
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
585 /* read the pixels */
986e461dc072 Initial revision
glantau
parents:
diff changeset
586 for(i=0;i<8;i++) {
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
587 pixels[0] = cm[block[0]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
588 pixels[1] = cm[block[1]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
589 pixels[2] = cm[block[2]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
590 pixels[3] = cm[block[3]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
591 pixels[4] = cm[block[4]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
592 pixels[5] = cm[block[5]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
593 pixels[6] = cm[block[6]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
594 pixels[7] = cm[block[7]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
595
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
596 pixels += line_size;
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
597 block += 8;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
598 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
599 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
600
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
601 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
602 int line_size)
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
603 {
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
604 int i;
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
605 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
606
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
607 /* read the pixels */
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
608 for(i=0;i<4;i++) {
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
609 pixels[0] = cm[block[0]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
610 pixels[1] = cm[block[1]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
611 pixels[2] = cm[block[2]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
612 pixels[3] = cm[block[3]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
613
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
614 pixels += line_size;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
615 block += 8;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
616 }
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
617 }
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
618
2257
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
619 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
620 int line_size)
2257
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
621 {
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
622 int i;
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
623 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
624
2257
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
625 /* read the pixels */
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
626 for(i=0;i<2;i++) {
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
627 pixels[0] = cm[block[0]];
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
628 pixels[1] = cm[block[1]];
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
629
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
630 pixels += line_size;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
631 block += 8;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
632 }
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
633 }
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
634
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
635 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
636 uint8_t *restrict pixels,
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
637 int line_size)
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
638 {
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
639 int i, j;
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
640
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
641 for (i = 0; i < 8; i++) {
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
642 for (j = 0; j < 8; j++) {
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
643 if (*block < -128)
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
644 *pixels = 0;
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
645 else if (*block > 127)
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
646 *pixels = 255;
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
647 else
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
648 *pixels = (uint8_t)(*block + 128);
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
649 block++;
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
650 pixels++;
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
651 }
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
652 pixels += (line_size - 8);
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
653 }
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
654 }
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
655
11231
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
656 static void put_pixels_nonclamped_c(const DCTELEM *block, uint8_t *restrict pixels,
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
657 int line_size)
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
658 {
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
659 int i;
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
660
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
661 /* read the pixels */
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
662 for(i=0;i<8;i++) {
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
663 pixels[0] = block[0];
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
664 pixels[1] = block[1];
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
665 pixels[2] = block[2];
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
666 pixels[3] = block[3];
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
667 pixels[4] = block[4];
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
668 pixels[5] = block[5];
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
669 pixels[6] = block[6];
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
670 pixels[7] = block[7];
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
671
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
672 pixels += line_size;
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
673 block += 8;
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
674 }
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
675 }
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
676
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
677 static void add_pixels_clamped_c(const DCTELEM *block, uint8_t *restrict pixels,
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
678 int line_size)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
679 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
680 int i;
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
681 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
682
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
683 /* read the pixels */
986e461dc072 Initial revision
glantau
parents:
diff changeset
684 for(i=0;i<8;i++) {
516
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
685 pixels[0] = cm[pixels[0] + block[0]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
686 pixels[1] = cm[pixels[1] + block[1]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
687 pixels[2] = cm[pixels[2] + block[2]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
688 pixels[3] = cm[pixels[3] + block[3]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
689 pixels[4] = cm[pixels[4] + block[4]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
690 pixels[5] = cm[pixels[5] + block[5]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
691 pixels[6] = cm[pixels[6] + block[6]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
692 pixels[7] = cm[pixels[7] + block[7]];
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
693 pixels += line_size;
0cb7ab35ccb2 * Detect "retrict" keyword
mellum
parents: 493
diff changeset
694 block += 8;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
695 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
696 }
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
697
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
698 static void add_pixels_clamped4_c(const DCTELEM *block, uint8_t *restrict pixels,
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
699 int line_size)
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
700 {
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
701 int i;
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
702 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
703
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
704 /* read the pixels */
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
705 for(i=0;i<4;i++) {
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
706 pixels[0] = cm[pixels[0] + block[0]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
707 pixels[1] = cm[pixels[1] + block[1]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
708 pixels[2] = cm[pixels[2] + block[2]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
709 pixels[3] = cm[pixels[3] + block[3]];
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
710 pixels += line_size;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
711 block += 8;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
712 }
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
713 }
2257
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
714
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
715 static void add_pixels_clamped2_c(const DCTELEM *block, uint8_t *restrict pixels,
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
716 int line_size)
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
717 {
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
718 int i;
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
719 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
720
2257
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
721 /* read the pixels */
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
722 for(i=0;i<2;i++) {
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
723 pixels[0] = cm[pixels[0] + block[0]];
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
724 pixels[1] = cm[pixels[1] + block[1]];
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
725 pixels += line_size;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
726 block += 8;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
727 }
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
728 }
2763
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
729
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
730 static void add_pixels8_c(uint8_t *restrict pixels, DCTELEM *block, int line_size)
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
731 {
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
732 int i;
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
733 for(i=0;i<8;i++) {
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
734 pixels[0] += block[0];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
735 pixels[1] += block[1];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
736 pixels[2] += block[2];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
737 pixels[3] += block[3];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
738 pixels[4] += block[4];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
739 pixels[5] += block[5];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
740 pixels[6] += block[6];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
741 pixels[7] += block[7];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
742 pixels += line_size;
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
743 block += 8;
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
744 }
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
745 }
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
746
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
747 static void add_pixels4_c(uint8_t *restrict pixels, DCTELEM *block, int line_size)
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
748 {
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
749 int i;
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
750 for(i=0;i<4;i++) {
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
751 pixels[0] += block[0];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
752 pixels[1] += block[1];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
753 pixels[2] += block[2];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
754 pixels[3] += block[3];
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
755 pixels += line_size;
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
756 block += 4;
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
757 }
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
758 }
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
759
4988
689490842cf5 factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents: 4749
diff changeset
760 static int sum_abs_dctelem_c(DCTELEM *block)
689490842cf5 factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents: 4749
diff changeset
761 {
689490842cf5 factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents: 4749
diff changeset
762 int sum=0, i;
689490842cf5 factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents: 4749
diff changeset
763 for(i=0; i<64; i++)
689490842cf5 factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents: 4749
diff changeset
764 sum+= FFABS(block[i]);
689490842cf5 factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents: 4749
diff changeset
765 return sum;
689490842cf5 factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents: 4749
diff changeset
766 }
689490842cf5 factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents: 4749
diff changeset
767
11231
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
768 static void fill_block16_c(uint8_t *block, uint8_t value, int line_size, int h)
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
769 {
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
770 int i;
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
771
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
772 for (i = 0; i < h; i++) {
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
773 memset(block, value, 16);
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
774 block += line_size;
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
775 }
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
776 }
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
777
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
778 static void fill_block8_c(uint8_t *block, uint8_t value, int line_size, int h)
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
779 {
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
780 int i;
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
781
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
782 for (i = 0; i < h; i++) {
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
783 memset(block, value, 8);
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
784 block += line_size;
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
785 }
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
786 }
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
787
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
788 static void scale_block_c(const uint8_t src[64]/*align 8*/, uint8_t *dst/*align 8*/, int linesize)
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
789 {
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
790 int i, j;
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
791 uint16_t *dst1 = dst;
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
792 uint16_t *dst2 = dst + linesize;
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
793
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
794 for (j = 0; j < 8; j++) {
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
795 for (i = 0; i < 8; i++) {
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
796 dst1[i] = dst2[i] = src[i] * 0x0101;
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
797 }
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
798 src += 8;
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
799 dst1 += linesize;
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
800 dst2 += linesize;
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
801 }
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
802 }
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
803
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
804 #if 0
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
805
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
806 #define PIXOP2(OPNAME, OP) \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
807 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
808 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
809 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
810 for(i=0; i<h; i++){\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
811 OP(*((uint64_t*)block), AV_RN64(pixels));\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
812 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
813 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
814 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
815 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
816 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
817 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
818 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
819 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
820 for(i=0; i<h; i++){\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
821 const uint64_t a= AV_RN64(pixels );\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
822 const uint64_t b= AV_RN64(pixels+1);\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
823 OP(*((uint64_t*)block), (a&b) + (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
824 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
825 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
826 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
827 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
828 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
829 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
830 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
831 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
832 for(i=0; i<h; i++){\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
833 const uint64_t a= AV_RN64(pixels );\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
834 const uint64_t b= AV_RN64(pixels+1);\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
835 OP(*((uint64_t*)block), (a|b) - (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
836 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
837 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
838 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
839 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
840 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
841 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
842 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
843 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
844 for(i=0; i<h; i++){\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
845 const uint64_t a= AV_RN64(pixels );\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
846 const uint64_t b= AV_RN64(pixels+line_size);\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
847 OP(*((uint64_t*)block), (a&b) + (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
848 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
849 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
850 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
851 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
852 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
853 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
854 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
855 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
856 for(i=0; i<h; i++){\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
857 const uint64_t a= AV_RN64(pixels );\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
858 const uint64_t b= AV_RN64(pixels+line_size);\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
859 OP(*((uint64_t*)block), (a|b) - (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
860 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
861 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
862 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
863 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
864 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
865 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
866 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
867 int i;\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
868 const uint64_t a= AV_RN64(pixels );\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
869 const uint64_t b= AV_RN64(pixels+1);\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
870 uint64_t l0= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
871 + (b&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
872 + 0x0202020202020202ULL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
873 uint64_t h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
874 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
875 uint64_t l1,h1;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
876 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
877 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
878 for(i=0; i<h; i+=2){\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
879 uint64_t a= AV_RN64(pixels );\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
880 uint64_t b= AV_RN64(pixels+1);\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
881 l1= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
882 + (b&0x0303030303030303ULL);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
883 h1= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
884 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
885 OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
886 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
887 block +=line_size;\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
888 a= AV_RN64(pixels );\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
889 b= AV_RN64(pixels+1);\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
890 l0= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
891 + (b&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
892 + 0x0202020202020202ULL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
893 h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
894 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
895 OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
896 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
897 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
898 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
899 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
900 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
901 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
902 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
903 int i;\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
904 const uint64_t a= AV_RN64(pixels );\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
905 const uint64_t b= AV_RN64(pixels+1);\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
906 uint64_t l0= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
907 + (b&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
908 + 0x0101010101010101ULL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
909 uint64_t h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
910 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
911 uint64_t l1,h1;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
912 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
913 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
914 for(i=0; i<h; i+=2){\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
915 uint64_t a= AV_RN64(pixels );\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
916 uint64_t b= AV_RN64(pixels+1);\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
917 l1= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
918 + (b&0x0303030303030303ULL);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
919 h1= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
920 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
921 OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
922 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
923 block +=line_size;\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
924 a= AV_RN64(pixels );\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
925 b= AV_RN64(pixels+1);\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
926 l0= (a&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
927 + (b&0x0303030303030303ULL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
928 + 0x0101010101010101ULL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
929 h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
930 + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
931 OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
932 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
933 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
934 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
935 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
936 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
937 CALL_2X_PIXELS(OPNAME ## _pixels16_c , OPNAME ## _pixels_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
938 CALL_2X_PIXELS(OPNAME ## _pixels16_x2_c , OPNAME ## _pixels_x2_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
939 CALL_2X_PIXELS(OPNAME ## _pixels16_y2_c , OPNAME ## _pixels_y2_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
940 CALL_2X_PIXELS(OPNAME ## _pixels16_xy2_c, OPNAME ## _pixels_xy2_c, 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
941 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
942 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
943 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
944
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
945 #define op_avg(a, b) a = ( ((a)|(b)) - ((((a)^(b))&0xFEFEFEFEFEFEFEFEULL)>>1) )
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
946 #else // 64 bit variant
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
947
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
948 #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
949 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
950 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
951 for(i=0; i<h; i++){\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
952 OP(*((uint16_t*)(block )), AV_RN16(pixels ));\
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
953 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
954 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
955 }\
85b71f9f7450 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 }\
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
957 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
958 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
959 for(i=0; i<h; i++){\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
960 OP(*((uint32_t*)(block )), AV_RN32(pixels ));\
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
961 pixels+=line_size;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
962 block +=line_size;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
963 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
964 }\
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
965 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
966 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
967 for(i=0; i<h; i++){\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
968 OP(*((uint32_t*)(block )), AV_RN32(pixels ));\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
969 OP(*((uint32_t*)(block+4)), AV_RN32(pixels+4));\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
970 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
971 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
972 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
973 }\
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
974 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
975 OPNAME ## _pixels8_c(block, pixels, line_size, h);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
976 }\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
977 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
978 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
979 int src_stride1, int src_stride2, int h){\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
980 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
981 for(i=0; i<h; i++){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
982 uint32_t a,b;\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
983 a= AV_RN32(&src1[i*src_stride1 ]);\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
984 b= AV_RN32(&src2[i*src_stride2 ]);\
1264
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
985 OP(*((uint32_t*)&dst[i*dst_stride ]), no_rnd_avg32(a, b));\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
986 a= AV_RN32(&src1[i*src_stride1+4]);\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
987 b= AV_RN32(&src2[i*src_stride2+4]);\
1264
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
988 OP(*((uint32_t*)&dst[i*dst_stride+4]), no_rnd_avg32(a, b));\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
989 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
990 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
991 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
992 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
993 int src_stride1, int src_stride2, int h){\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
994 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
995 for(i=0; i<h; i++){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
996 uint32_t a,b;\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
997 a= AV_RN32(&src1[i*src_stride1 ]);\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
998 b= AV_RN32(&src2[i*src_stride2 ]);\
1264
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
999 OP(*((uint32_t*)&dst[i*dst_stride ]), rnd_avg32(a, b));\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1000 a= AV_RN32(&src1[i*src_stride1+4]);\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1001 b= AV_RN32(&src2[i*src_stride2+4]);\
1264
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
1002 OP(*((uint32_t*)&dst[i*dst_stride+4]), rnd_avg32(a, b));\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1003 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1004 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1005 \
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1006 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
1007 int src_stride1, int src_stride2, int h){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1008 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1009 for(i=0; i<h; i++){\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1010 uint32_t a,b;\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1011 a= AV_RN32(&src1[i*src_stride1 ]);\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1012 b= AV_RN32(&src2[i*src_stride2 ]);\
1264
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
1013 OP(*((uint32_t*)&dst[i*dst_stride ]), rnd_avg32(a, b));\
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1014 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1015 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1016 \
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
1017 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
1018 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
1019 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
1020 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
1021 uint32_t a,b;\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1022 a= AV_RN16(&src1[i*src_stride1 ]);\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1023 b= AV_RN16(&src2[i*src_stride2 ]);\
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
1024 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
1025 }\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1026 }\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1027 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1028 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
1029 int src_stride1, int src_stride2, int h){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1030 OPNAME ## _pixels8_l2(dst , src1 , src2 , dst_stride, src_stride1, src_stride2, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1031 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
1032 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1033 \
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1034 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
1035 int src_stride1, int src_stride2, int h){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1036 OPNAME ## _no_rnd_pixels8_l2(dst , src1 , src2 , dst_stride, src_stride1, src_stride2, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1037 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
1038 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1039 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1040 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
1041 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
1042 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1043 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1044 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
1045 OPNAME ## _pixels8_l2(block, pixels, pixels+1, line_size, line_size, line_size, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1046 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1047 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1048 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
1049 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
1050 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1051 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1052 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
1053 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
1054 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1055 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1056 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
1057 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
1058 int i;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1059 for(i=0; i<h; i++){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1060 uint32_t a, b, c, d, l0, l1, h0, h1;\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1061 a= AV_RN32(&src1[i*src_stride1]);\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1062 b= AV_RN32(&src2[i*src_stride2]);\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1063 c= AV_RN32(&src3[i*src_stride3]);\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1064 d= AV_RN32(&src4[i*src_stride4]);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1065 l0= (a&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1066 + (b&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1067 + 0x02020202UL;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1068 h0= ((a&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1069 + ((b&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1070 l1= (c&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1071 + (d&0x03030303UL);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1072 h1= ((c&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1073 + ((d&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1074 OP(*((uint32_t*)&dst[i*dst_stride]), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1075 a= AV_RN32(&src1[i*src_stride1+4]);\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1076 b= AV_RN32(&src2[i*src_stride2+4]);\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1077 c= AV_RN32(&src3[i*src_stride3+4]);\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1078 d= AV_RN32(&src4[i*src_stride4+4]);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1079 l0= (a&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1080 + (b&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1081 + 0x02020202UL;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1082 h0= ((a&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1083 + ((b&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1084 l1= (c&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1085 + (d&0x03030303UL);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1086 h1= ((c&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1087 + ((d&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1088 OP(*((uint32_t*)&dst[i*dst_stride+4]), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1089 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1090 }\
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
1091 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1092 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
1093 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
1094 }\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1095 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1096 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
1097 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
1098 }\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1099 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1100 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
1101 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
1102 }\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1103 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1104 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
1105 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
1106 }\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1107 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1108 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
1109 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
1110 int i;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1111 for(i=0; i<h; i++){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1112 uint32_t a, b, c, d, l0, l1, h0, h1;\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1113 a= AV_RN32(&src1[i*src_stride1]);\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1114 b= AV_RN32(&src2[i*src_stride2]);\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1115 c= AV_RN32(&src3[i*src_stride3]);\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1116 d= AV_RN32(&src4[i*src_stride4]);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1117 l0= (a&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1118 + (b&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1119 + 0x01010101UL;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1120 h0= ((a&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1121 + ((b&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1122 l1= (c&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1123 + (d&0x03030303UL);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1124 h1= ((c&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1125 + ((d&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1126 OP(*((uint32_t*)&dst[i*dst_stride]), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1127 a= AV_RN32(&src1[i*src_stride1+4]);\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1128 b= AV_RN32(&src2[i*src_stride2+4]);\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1129 c= AV_RN32(&src3[i*src_stride3+4]);\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1130 d= AV_RN32(&src4[i*src_stride4+4]);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1131 l0= (a&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1132 + (b&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1133 + 0x01010101UL;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1134 h0= ((a&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1135 + ((b&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1136 l1= (c&0x03030303UL)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1137 + (d&0x03030303UL);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1138 h1= ((c&0xFCFCFCFCUL)>>2)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1139 + ((d&0xFCFCFCFCUL)>>2);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1140 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
1141 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1142 }\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1143 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
1144 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
1145 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
1146 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
1147 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1148 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
1149 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
1150 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
1151 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
1152 }\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1153 \
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
1154 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
1155 {\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1156 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
1157 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
1158 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
1159 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
1160 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
1161 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1162 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
1163 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
1164 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
1165 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
1166 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
1167 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
1168 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1169 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
1170 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
1171 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1172 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
1173 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
1174 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1175 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
1176 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
1177 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
1178 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
1179 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1180 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
1181 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
1182 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
1183 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
1184 }\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1185 }\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1186 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1187 static inline void 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
1188 {\
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1189 int i;\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1190 const uint32_t a= AV_RN32(pixels );\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1191 const uint32_t b= AV_RN32(pixels+1);\
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
1192 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
1193 + (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
1194 + 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
1195 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
1196 + ((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
1197 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
1198 \
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1199 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
1200 for(i=0; i<h; i+=2){\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1201 uint32_t a= AV_RN32(pixels );\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1202 uint32_t b= AV_RN32(pixels+1);\
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 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
1204 + (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
1205 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
1206 + ((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
1207 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
1208 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
1209 block +=line_size;\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1210 a= AV_RN32(pixels );\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1211 b= AV_RN32(pixels+1);\
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
1212 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
1213 + (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
1214 + 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
1215 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
1216 + ((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
1217 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
1218 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
1219 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
1220 }\
85b71f9f7450 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 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1223 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
1224 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1225 int j;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1226 for(j=0; j<2; j++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1227 int i;\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1228 const uint32_t a= AV_RN32(pixels );\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1229 const uint32_t b= AV_RN32(pixels+1);\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1230 uint32_t l0= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1231 + (b&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1232 + 0x02020202UL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1233 uint32_t h0= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1234 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1235 uint32_t l1,h1;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1236 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1237 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1238 for(i=0; i<h; i+=2){\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1239 uint32_t a= AV_RN32(pixels );\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1240 uint32_t b= AV_RN32(pixels+1);\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1241 l1= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1242 + (b&0x03030303UL);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1243 h1= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1244 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1245 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1246 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1247 block +=line_size;\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1248 a= AV_RN32(pixels );\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1249 b= AV_RN32(pixels+1);\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1250 l0= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1251 + (b&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1252 + 0x02020202UL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1253 h0= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1254 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1255 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1256 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1257 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1258 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1259 pixels+=4-line_size*(h+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1260 block +=4-line_size*h;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1261 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1262 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1263 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1264 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
1265 {\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1266 int j;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1267 for(j=0; j<2; j++){\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1268 int i;\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1269 const uint32_t a= AV_RN32(pixels );\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1270 const uint32_t b= AV_RN32(pixels+1);\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1271 uint32_t l0= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1272 + (b&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1273 + 0x01010101UL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1274 uint32_t h0= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1275 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1276 uint32_t l1,h1;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1277 \
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1278 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1279 for(i=0; i<h; i+=2){\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1280 uint32_t a= AV_RN32(pixels );\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1281 uint32_t b= AV_RN32(pixels+1);\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1282 l1= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1283 + (b&0x03030303UL);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1284 h1= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1285 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1286 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1287 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1288 block +=line_size;\
5520
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1289 a= AV_RN32(pixels );\
c16a59ef6a86 * renaming (ST|LD)(16|32|64) -> AV_(R|W)N(16|32|64)
romansh
parents: 5411
diff changeset
1290 b= AV_RN32(pixels+1);\
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1291 l0= (a&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1292 + (b&0x03030303UL)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1293 + 0x01010101UL;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1294 h0= ((a&0xFCFCFCFCUL)>>2)\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1295 + ((b&0xFCFCFCFCUL)>>2);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1296 OP(*((uint32_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0FUL));\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1297 pixels+=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1298 block +=line_size;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1299 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1300 pixels+=4-line_size*(h+1);\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1301 block +=4-line_size*h;\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1302 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1303 }\
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1304 \
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1305 CALL_2X_PIXELS(OPNAME ## _pixels16_c , OPNAME ## _pixels8_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1306 CALL_2X_PIXELS(OPNAME ## _pixels16_x2_c , OPNAME ## _pixels8_x2_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1307 CALL_2X_PIXELS(OPNAME ## _pixels16_y2_c , OPNAME ## _pixels8_y2_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1308 CALL_2X_PIXELS(OPNAME ## _pixels16_xy2_c, OPNAME ## _pixels8_xy2_c, 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1309 CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_c , OPNAME ## _pixels8_c , 8)\
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
1310 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
1311 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
1312 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
1313
1264
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
1314 #define op_avg(a, b) a = rnd_avg32(a, b)
385
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1315 #endif
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1316 #define op_put(a, b) a = b
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1317
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1318 PIXOP2(avg, op_avg)
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1319 PIXOP2(put, op_put)
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1320 #undef op_avg
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1321 #undef op_put
7ac7a48fbe5e new hopefully faster MC
michaelni
parents: 324
diff changeset
1322
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
1323 #define avg2(a,b) ((a+b+1)>>1)
986e461dc072 Initial revision
glantau
parents:
diff changeset
1324 #define avg4(a,b,c,d) ((a+b+c+d+2)>>2)
986e461dc072 Initial revision
glantau
parents:
diff changeset
1325
1864
9a73e6dc5d68 fix motion compensation with (x+1/2,y+1/2) MVs
michael
parents: 1784
diff changeset
1326 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
1327 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
1328 }
9a73e6dc5d68 fix motion compensation with (x+1/2,y+1/2) MVs
michael
parents: 1784
diff changeset
1329
9a73e6dc5d68 fix motion compensation with (x+1/2,y+1/2) MVs
michael
parents: 1784
diff changeset
1330 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
1331 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
1332 }
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1333
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1334 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
1335 {
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1336 const int A=(16-x16)*(16-y16);
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1337 const int B=( x16)*(16-y16);
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1338 const int C=(16-x16)*( y16);
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1339 const int D=( x16)*( y16);
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1340 int i;
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1341
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1342 for(i=0; i<h; i++)
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1343 {
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1344 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
1345 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
1346 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
1347 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
1348 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
1349 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
1350 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
1351 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
1352 dst+= stride;
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1353 src+= stride;
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1354 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1355 }
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1356
3248
7aa9f80e7954 mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents: 3245
diff changeset
1357 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
1358 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
1359 {
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1360 int y, vx, vy;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1361 const int s= 1<<shift;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
1362
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1363 width--;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1364 height--;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1365
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1366 for(y=0; y<h; y++){
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1367 int x;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1368
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1369 vx= ox;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1370 vy= oy;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1371 for(x=0; x<8; x++){ //XXX FIXME optimize
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1372 int src_x, src_y, frac_x, frac_y, index;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1373
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1374 src_x= vx>>16;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1375 src_y= vy>>16;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1376 frac_x= src_x&(s-1);
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1377 frac_y= src_y&(s-1);
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1378 src_x>>=shift;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1379 src_y>>=shift;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
1380
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1381 if((unsigned)src_x < width){
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1382 if((unsigned)src_y < height){
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1383 index= src_x + src_y*stride;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1384 dst[y*stride + x]= ( ( src[index ]*(s-frac_x)
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1385 + src[index +1]* frac_x )*(s-frac_y)
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1386 + ( src[index+stride ]*(s-frac_x)
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1387 + src[index+stride+1]* frac_x )* frac_y
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1388 + r)>>(shift*2);
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1389 }else{
4594
a96d905dcbaa Add av_ prefix to clip functions
reimar
parents: 4296
diff changeset
1390 index= src_x + av_clip(src_y, 0, height)*stride;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
1391 dst[y*stride + x]= ( ( src[index ]*(s-frac_x)
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1392 + src[index +1]* frac_x )*s
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1393 + r)>>(shift*2);
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1394 }
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1395 }else{
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1396 if((unsigned)src_y < height){
4594
a96d905dcbaa Add av_ prefix to clip functions
reimar
parents: 4296
diff changeset
1397 index= av_clip(src_x, 0, width) + src_y*stride;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
1398 dst[y*stride + x]= ( ( src[index ]*(s-frac_y)
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1399 + src[index+stride ]* frac_y )*s
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1400 + r)>>(shift*2);
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1401 }else{
4594
a96d905dcbaa Add av_ prefix to clip functions
reimar
parents: 4296
diff changeset
1402 index= av_clip(src_x, 0, width) + av_clip(src_y, 0, height)*stride;
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1403 dst[y*stride + x]= src[index ];
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1404 }
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1405 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
1406
753
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1407 vx+= dxx;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1408 vy+= dyx;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1409 }
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1410 ox += dxy;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1411 oy += dyy;
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1412 }
8e1f0939d15d complete mpeg4 GMC decoding support
michaelni
parents: 751
diff changeset
1413 }
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
1414
85b71f9f7450 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 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
1416 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
1417 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
1418 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
1419 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
1420 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
1421 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1422 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1423
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1424 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
1425 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
1426 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
1427 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1428 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
1429 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1430 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
1431 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
1432 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1433 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1434
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1435 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
1436 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
1437 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
1438 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1439 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
1440 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1441 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
1442 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
1443 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1444 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
1445
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
1446 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
1447 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
1448 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
1449 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1450 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
1451 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1452 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
1453 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
1454 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1455 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
1456
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
1457 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
1458 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
1459 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
1460 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1461 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
1462 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1463 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
1464 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
1465 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1466 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1467
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1468 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
1469 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
1470 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
1471 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1472 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
1473 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1474 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
1475 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
1476 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1477 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1478
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1479 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
1480 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
1481 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
1482 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1483 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
1484 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1485 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
1486 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
1487 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1488 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1489
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1490 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
1491 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
1492 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
1493 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1494 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
1495 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1496 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
1497 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
1498 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1499 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1500
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1501 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
1502 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
1503 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
1504 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1505 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
1506 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1507 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
1508 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
1509 }
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
1510 }
1319
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1511
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1512 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
1513 switch(width){
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1514 case 2: avg_pixels2_c (dst, src, stride, height); break;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1515 case 4: avg_pixels4_c (dst, src, stride, height); break;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1516 case 8: avg_pixels8_c (dst, src, stride, height); break;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1517 case 16:avg_pixels16_c(dst, src, stride, height); break;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1518 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1519 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1520
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1521 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
1522 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1523 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1524 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1525 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
1526 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1527 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1528 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1529 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1530 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1531
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1532 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
1533 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1534 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1535 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1536 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
1537 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1538 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1539 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1540 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1541 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
1542
1319
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1543 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
1544 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1545 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1546 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1547 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
1548 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1549 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1550 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1551 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1552 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
1553
1319
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1554 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
1555 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1556 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1557 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1558 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
1559 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1560 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1561 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1562 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1563 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1564
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1565 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
1566 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1567 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1568 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1569 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
1570 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1571 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1572 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1573 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1574 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1575
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1576 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
1577 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1578 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1579 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1580 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
1581 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1582 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1583 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1584 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1585 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1586
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1587 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
1588 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1589 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1590 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1591 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
1592 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1593 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1594 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1595 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1596 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1597
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1598 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
1599 int i,j;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1600 for (i=0; i < height; i++) {
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1601 for (j=0; j < width; j++) {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
1602 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
1603 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1604 src += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1605 dst += stride;
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1606 }
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
1607 }
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
1608 #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
1609 #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
1610 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
1611 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
1612 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
1613 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
1614 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
1615 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
1616 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
1617 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
1618 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
1619 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
1620 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
1621 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
1622 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
1623 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
1624 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
1625 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
1626 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
1627 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
1628 #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
1629
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1630 #define H264_CHROMA_MC(OPNAME, OP)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1631 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
1632 const int A=(8-x)*(8-y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1633 const int B=( x)*(8-y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1634 const int C=(8-x)*( y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1635 const int D=( x)*( y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1636 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1637 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1638 assert(x<8 && y<8 && x>=0 && y>=0);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1639 \
6052
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1640 if(D){\
6054
de71ee8112f4 More consistent { placement.
michael
parents: 6053
diff changeset
1641 for(i=0; i<h; i++){\
6053
michael
parents: 6052
diff changeset
1642 OP(dst[0], (A*src[0] + B*src[1] + C*src[stride+0] + D*src[stride+1]));\
michael
parents: 6052
diff changeset
1643 OP(dst[1], (A*src[1] + B*src[2] + C*src[stride+1] + D*src[stride+2]));\
michael
parents: 6052
diff changeset
1644 dst+= stride;\
michael
parents: 6052
diff changeset
1645 src+= stride;\
michael
parents: 6052
diff changeset
1646 }\
6052
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1647 }else{\
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1648 const int E= B+C;\
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1649 const int step= C ? stride : 1;\
6054
de71ee8112f4 More consistent { placement.
michael
parents: 6053
diff changeset
1650 for(i=0; i<h; i++){\
6052
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1651 OP(dst[0], (A*src[0] + E*src[step+0]));\
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1652 OP(dst[1], (A*src[1] + E*src[step+1]));\
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1653 dst+= stride;\
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1654 src+= stride;\
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1655 }\
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1656 }\
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1657 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1658 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1659 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
1660 const int A=(8-x)*(8-y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1661 const int B=( x)*(8-y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1662 const int C=(8-x)*( y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1663 const int D=( x)*( y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1664 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1665 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1666 assert(x<8 && y<8 && x>=0 && y>=0);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1667 \
6052
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1668 if(D){\
6054
de71ee8112f4 More consistent { placement.
michael
parents: 6053
diff changeset
1669 for(i=0; i<h; i++){\
6053
michael
parents: 6052
diff changeset
1670 OP(dst[0], (A*src[0] + B*src[1] + C*src[stride+0] + D*src[stride+1]));\
michael
parents: 6052
diff changeset
1671 OP(dst[1], (A*src[1] + B*src[2] + C*src[stride+1] + D*src[stride+2]));\
michael
parents: 6052
diff changeset
1672 OP(dst[2], (A*src[2] + B*src[3] + C*src[stride+2] + D*src[stride+3]));\
michael
parents: 6052
diff changeset
1673 OP(dst[3], (A*src[3] + B*src[4] + C*src[stride+3] + D*src[stride+4]));\
michael
parents: 6052
diff changeset
1674 dst+= stride;\
michael
parents: 6052
diff changeset
1675 src+= stride;\
michael
parents: 6052
diff changeset
1676 }\
6052
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1677 }else{\
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1678 const int E= B+C;\
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1679 const int step= C ? stride : 1;\
6054
de71ee8112f4 More consistent { placement.
michael
parents: 6053
diff changeset
1680 for(i=0; i<h; i++){\
6052
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1681 OP(dst[0], (A*src[0] + E*src[step+0]));\
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1682 OP(dst[1], (A*src[1] + E*src[step+1]));\
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1683 OP(dst[2], (A*src[2] + E*src[step+2]));\
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1684 OP(dst[3], (A*src[3] + E*src[step+3]));\
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1685 dst+= stride;\
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1686 src+= stride;\
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1687 }\
c90798ac28ee ~15% faster h264_chroma_mc2/4_c() these also prevent some possible out
michael
parents: 6051
diff changeset
1688 }\
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1689 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1690 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1691 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
1692 const int A=(8-x)*(8-y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1693 const int B=( x)*(8-y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1694 const int C=(8-x)*( y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1695 const int D=( x)*( y);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1696 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1697 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1698 assert(x<8 && y<8 && x>=0 && y>=0);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1699 \
6051
1e3b5597505a 30% faster h264_chroma_mc8_c(), this also prevents a possible out of
michael
parents: 6001
diff changeset
1700 if(D){\
6054
de71ee8112f4 More consistent { placement.
michael
parents: 6053
diff changeset
1701 for(i=0; i<h; i++){\
6053
michael
parents: 6052
diff changeset
1702 OP(dst[0], (A*src[0] + B*src[1] + C*src[stride+0] + D*src[stride+1]));\
michael
parents: 6052
diff changeset
1703 OP(dst[1], (A*src[1] + B*src[2] + C*src[stride+1] + D*src[stride+2]));\
michael
parents: 6052
diff changeset
1704 OP(dst[2], (A*src[2] + B*src[3] + C*src[stride+2] + D*src[stride+3]));\
michael
parents: 6052
diff changeset
1705 OP(dst[3], (A*src[3] + B*src[4] + C*src[stride+3] + D*src[stride+4]));\
michael
parents: 6052
diff changeset
1706 OP(dst[4], (A*src[4] + B*src[5] + C*src[stride+4] + D*src[stride+5]));\
michael
parents: 6052
diff changeset
1707 OP(dst[5], (A*src[5] + B*src[6] + C*src[stride+5] + D*src[stride+6]));\
michael
parents: 6052
diff changeset
1708 OP(dst[6], (A*src[6] + B*src[7] + C*src[stride+6] + D*src[stride+7]));\
michael
parents: 6052
diff changeset
1709 OP(dst[7], (A*src[7] + B*src[8] + C*src[stride+7] + D*src[stride+8]));\
michael
parents: 6052
diff changeset
1710 dst+= stride;\
michael
parents: 6052
diff changeset
1711 src+= stride;\
michael
parents: 6052
diff changeset
1712 }\
6051
1e3b5597505a 30% faster h264_chroma_mc8_c(), this also prevents a possible out of
michael
parents: 6001
diff changeset
1713 }else{\
1e3b5597505a 30% faster h264_chroma_mc8_c(), this also prevents a possible out of
michael
parents: 6001
diff changeset
1714 const int E= B+C;\
1e3b5597505a 30% faster h264_chroma_mc8_c(), this also prevents a possible out of
michael
parents: 6001
diff changeset
1715 const int step= C ? stride : 1;\
6054
de71ee8112f4 More consistent { placement.
michael
parents: 6053
diff changeset
1716 for(i=0; i<h; i++){\
6051
1e3b5597505a 30% faster h264_chroma_mc8_c(), this also prevents a possible out of
michael
parents: 6001
diff changeset
1717 OP(dst[0], (A*src[0] + E*src[step+0]));\
1e3b5597505a 30% faster h264_chroma_mc8_c(), this also prevents a possible out of
michael
parents: 6001
diff changeset
1718 OP(dst[1], (A*src[1] + E*src[step+1]));\
1e3b5597505a 30% faster h264_chroma_mc8_c(), this also prevents a possible out of
michael
parents: 6001
diff changeset
1719 OP(dst[2], (A*src[2] + E*src[step+2]));\
1e3b5597505a 30% faster h264_chroma_mc8_c(), this also prevents a possible out of
michael
parents: 6001
diff changeset
1720 OP(dst[3], (A*src[3] + E*src[step+3]));\
1e3b5597505a 30% faster h264_chroma_mc8_c(), this also prevents a possible out of
michael
parents: 6001
diff changeset
1721 OP(dst[4], (A*src[4] + E*src[step+4]));\
1e3b5597505a 30% faster h264_chroma_mc8_c(), this also prevents a possible out of
michael
parents: 6001
diff changeset
1722 OP(dst[5], (A*src[5] + E*src[step+5]));\
1e3b5597505a 30% faster h264_chroma_mc8_c(), this also prevents a possible out of
michael
parents: 6001
diff changeset
1723 OP(dst[6], (A*src[6] + E*src[step+6]));\
1e3b5597505a 30% faster h264_chroma_mc8_c(), this also prevents a possible out of
michael
parents: 6001
diff changeset
1724 OP(dst[7], (A*src[7] + E*src[step+7]));\
1e3b5597505a 30% faster h264_chroma_mc8_c(), this also prevents a possible out of
michael
parents: 6001
diff changeset
1725 dst+= stride;\
1e3b5597505a 30% faster h264_chroma_mc8_c(), this also prevents a possible out of
michael
parents: 6001
diff changeset
1726 src+= stride;\
1e3b5597505a 30% faster h264_chroma_mc8_c(), this also prevents a possible out of
michael
parents: 6001
diff changeset
1727 }\
1e3b5597505a 30% faster h264_chroma_mc8_c(), this also prevents a possible out of
michael
parents: 6001
diff changeset
1728 }\
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1729 }
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1730
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1731 #define op_avg(a, b) a = (((a)+(((b) + 32)>>6)+1)>>1)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1732 #define op_put(a, b) a = (((b) + 32)>>6)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1733
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1734 H264_CHROMA_MC(put_ , op_put)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1735 H264_CHROMA_MC(avg_ , op_avg)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1736 #undef op_avg
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1737 #undef op_put
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
1738
9439
ef3a7b711cc0 Rename put_no_rnd_h264_chroma* to reflect its usage in VC1 only
conrad
parents: 9437
diff changeset
1739 static void put_no_rnd_vc1_chroma_mc8_c(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y){
3663
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1740 const int A=(8-x)*(8-y);
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1741 const int B=( x)*(8-y);
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1742 const int C=(8-x)*( y);
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1743 const int D=( x)*( y);
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1744 int i;
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1745
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1746 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
1747
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1748 for(i=0; i<h; i++)
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1749 {
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1750 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
1751 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
1752 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
1753 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
1754 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
1755 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
1756 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
1757 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
1758 dst+= stride;
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1759 src+= stride;
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1760 }
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1761 }
8e180a3276fe New function for chroma MC (will be used in VC-1)
kostya
parents: 3656
diff changeset
1762
9440
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1763 static void avg_no_rnd_vc1_chroma_mc8_c(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int x, int y){
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1764 const int A=(8-x)*(8-y);
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1765 const int B=( x)*(8-y);
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1766 const int C=(8-x)*( y);
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1767 const int D=( x)*( y);
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1768 int i;
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1769
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1770 assert(x<8 && y<8 && x>=0 && y>=0);
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1771
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1772 for(i=0; i<h; i++)
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1773 {
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1774 dst[0] = avg2(dst[0], ((A*src[0] + B*src[1] + C*src[stride+0] + D*src[stride+1] + 32 - 4) >> 6));
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1775 dst[1] = avg2(dst[1], ((A*src[1] + B*src[2] + C*src[stride+1] + D*src[stride+2] + 32 - 4) >> 6));
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1776 dst[2] = avg2(dst[2], ((A*src[2] + B*src[3] + C*src[stride+2] + D*src[stride+3] + 32 - 4) >> 6));
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1777 dst[3] = avg2(dst[3], ((A*src[3] + B*src[4] + C*src[stride+3] + D*src[stride+4] + 32 - 4) >> 6));
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1778 dst[4] = avg2(dst[4], ((A*src[4] + B*src[5] + C*src[stride+4] + D*src[stride+5] + 32 - 4) >> 6));
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1779 dst[5] = avg2(dst[5], ((A*src[5] + B*src[6] + C*src[stride+5] + D*src[stride+6] + 32 - 4) >> 6));
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1780 dst[6] = avg2(dst[6], ((A*src[6] + B*src[7] + C*src[stride+6] + D*src[stride+7] + 32 - 4) >> 6));
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1781 dst[7] = avg2(dst[7], ((A*src[7] + B*src[8] + C*src[stride+7] + D*src[stride+8] + 32 - 4) >> 6));
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1782 dst+= stride;
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1783 src+= stride;
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1784 }
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1785 }
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
1786
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1787 #define QPEL_MC(r, OPNAME, RND, OP) \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1788 static void OPNAME ## mpeg4_qpel8_h_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
1789 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1790 int i;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1791 for(i=0; i<h; i++)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1792 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1793 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
1794 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
1795 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
1796 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
1797 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
1798 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
1799 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
1800 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
1801 dst+=dstStride;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1802 src+=srcStride;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1803 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1804 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1805 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1806 static void OPNAME ## mpeg4_qpel8_v_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1807 const int w=8;\
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
1808 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1809 int i;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1810 for(i=0; i<w; i++)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1811 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1812 const int src0= src[0*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1813 const int src1= src[1*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1814 const int src2= src[2*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1815 const int src3= src[3*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1816 const int src4= src[4*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1817 const int src5= src[5*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1818 const int src6= src[6*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1819 const int src7= src[7*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1820 const int src8= src[8*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1821 OP(dst[0*dstStride], (src0+src1)*20 - (src0+src2)*6 + (src1+src3)*3 - (src2+src4));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1822 OP(dst[1*dstStride], (src1+src2)*20 - (src0+src3)*6 + (src0+src4)*3 - (src1+src5));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1823 OP(dst[2*dstStride], (src2+src3)*20 - (src1+src4)*6 + (src0+src5)*3 - (src0+src6));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1824 OP(dst[3*dstStride], (src3+src4)*20 - (src2+src5)*6 + (src1+src6)*3 - (src0+src7));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1825 OP(dst[4*dstStride], (src4+src5)*20 - (src3+src6)*6 + (src2+src7)*3 - (src1+src8));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1826 OP(dst[5*dstStride], (src5+src6)*20 - (src4+src7)*6 + (src3+src8)*3 - (src2+src8));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1827 OP(dst[6*dstStride], (src6+src7)*20 - (src5+src8)*6 + (src4+src8)*3 - (src3+src7));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1828 OP(dst[7*dstStride], (src7+src8)*20 - (src6+src8)*6 + (src5+src7)*3 - (src4+src6));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1829 dst++;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1830 src++;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1831 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1832 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1833 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1834 static void OPNAME ## mpeg4_qpel16_h_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
1835 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1836 int i;\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1837 \
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1838 for(i=0; i<h; i++)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1839 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1840 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
1841 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
1842 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
1843 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
1844 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
1845 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
1846 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
1847 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
1848 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
1849 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
1850 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
1851 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
1852 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
1853 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
1854 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
1855 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
1856 dst+=dstStride;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1857 src+=srcStride;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1858 }\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1859 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1860 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1861 static void OPNAME ## mpeg4_qpel16_v_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
1862 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1863 int i;\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
1864 const int w=16;\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1865 for(i=0; i<w; i++)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1866 {\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1867 const int src0= src[0*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1868 const int src1= src[1*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1869 const int src2= src[2*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1870 const int src3= src[3*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1871 const int src4= src[4*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1872 const int src5= src[5*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1873 const int src6= src[6*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1874 const int src7= src[7*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1875 const int src8= src[8*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1876 const int src9= src[9*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1877 const int src10= src[10*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1878 const int src11= src[11*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1879 const int src12= src[12*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1880 const int src13= src[13*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1881 const int src14= src[14*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1882 const int src15= src[15*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1883 const int src16= src[16*srcStride];\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1884 OP(dst[ 0*dstStride], (src0 +src1 )*20 - (src0 +src2 )*6 + (src1 +src3 )*3 - (src2 +src4 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1885 OP(dst[ 1*dstStride], (src1 +src2 )*20 - (src0 +src3 )*6 + (src0 +src4 )*3 - (src1 +src5 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1886 OP(dst[ 2*dstStride], (src2 +src3 )*20 - (src1 +src4 )*6 + (src0 +src5 )*3 - (src0 +src6 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1887 OP(dst[ 3*dstStride], (src3 +src4 )*20 - (src2 +src5 )*6 + (src1 +src6 )*3 - (src0 +src7 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1888 OP(dst[ 4*dstStride], (src4 +src5 )*20 - (src3 +src6 )*6 + (src2 +src7 )*3 - (src1 +src8 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1889 OP(dst[ 5*dstStride], (src5 +src6 )*20 - (src4 +src7 )*6 + (src3 +src8 )*3 - (src2 +src9 ));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1890 OP(dst[ 6*dstStride], (src6 +src7 )*20 - (src5 +src8 )*6 + (src4 +src9 )*3 - (src3 +src10));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1891 OP(dst[ 7*dstStride], (src7 +src8 )*20 - (src6 +src9 )*6 + (src5 +src10)*3 - (src4 +src11));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1892 OP(dst[ 8*dstStride], (src8 +src9 )*20 - (src7 +src10)*6 + (src6 +src11)*3 - (src5 +src12));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1893 OP(dst[ 9*dstStride], (src9 +src10)*20 - (src8 +src11)*6 + (src7 +src12)*3 - (src6 +src13));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1894 OP(dst[10*dstStride], (src10+src11)*20 - (src9 +src12)*6 + (src8 +src13)*3 - (src7 +src14));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1895 OP(dst[11*dstStride], (src11+src12)*20 - (src10+src13)*6 + (src9 +src14)*3 - (src8 +src15));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1896 OP(dst[12*dstStride], (src12+src13)*20 - (src11+src14)*6 + (src10+src15)*3 - (src9 +src16));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1897 OP(dst[13*dstStride], (src13+src14)*20 - (src12+src15)*6 + (src11+src16)*3 - (src10+src16));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1898 OP(dst[14*dstStride], (src14+src15)*20 - (src13+src16)*6 + (src12+src16)*3 - (src11+src15));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1899 OP(dst[15*dstStride], (src15+src16)*20 - (src14+src16)*6 + (src13+src15)*3 - (src12+src14));\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1900 dst++;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1901 src++;\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1902 }\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1903 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1904 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1905 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
1906 OPNAME ## pixels8_c(dst, src, stride, 8);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1907 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1908 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1909 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
1910 uint8_t half[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1911 put ## RND ## mpeg4_qpel8_h_lowpass(half, src, 8, stride, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1912 OPNAME ## pixels8_l2(dst, src, half, stride, stride, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1913 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1914 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1915 static void OPNAME ## qpel8_mc20_c(uint8_t *dst, uint8_t *src, int stride){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1916 OPNAME ## mpeg4_qpel8_h_lowpass(dst, src, stride, stride, 8);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1917 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1918 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1919 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
1920 uint8_t half[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1921 put ## RND ## mpeg4_qpel8_h_lowpass(half, src, 8, stride, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1922 OPNAME ## pixels8_l2(dst, src+1, half, stride, stride, 8, 8);\
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 ## qpel8_mc01_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[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1927 uint8_t half[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1928 copy_block9(full, src, 16, stride, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1929 put ## RND ## mpeg4_qpel8_v_lowpass(half, full, 8, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1930 OPNAME ## pixels8_l2(dst, full, half, stride, 16, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1931 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1932 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1933 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
1934 uint8_t full[16*9];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1935 copy_block9(full, src, 16, stride, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1936 OPNAME ## mpeg4_qpel8_v_lowpass(dst, full, stride, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1937 }\
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1938 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1939 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
1940 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1941 uint8_t half[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1942 copy_block9(full, src, 16, stride, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1943 put ## RND ## mpeg4_qpel8_v_lowpass(half, full, 8, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1944 OPNAME ## pixels8_l2(dst, full+16, half, stride, 16, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1945 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1946 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
1947 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1948 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1949 uint8_t halfV[64];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1950 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1951 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1952 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1953 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full, 8, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1954 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1955 OPNAME ## pixels8_l4(dst, full, halfH, halfV, halfHV, stride, 16, 8, 8, 8, 8);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
1956 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1957 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
1958 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1959 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1960 uint8_t halfHV[64];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1961 copy_block9(full, src, 16, stride, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1962 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1963 put ## RND ## pixels8_l2(halfH, halfH, full, 8, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1964 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1965 OPNAME ## pixels8_l2(dst, halfH, halfHV, stride, 8, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1966 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1967 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
1968 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1969 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1970 uint8_t halfV[64];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1971 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1972 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1973 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1974 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full+1, 8, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1975 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1976 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
1977 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1978 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
1979 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1980 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1981 uint8_t halfHV[64];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1982 copy_block9(full, src, 16, stride, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1983 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1984 put ## RND ## pixels8_l2(halfH, halfH, full+1, 8, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1985 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1986 OPNAME ## pixels8_l2(dst, halfH, halfHV, stride, 8, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1987 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1988 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
1989 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1990 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1991 uint8_t halfV[64];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1992 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1993 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1994 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1995 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full, 8, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
1996 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
1997 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
1998 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
1999 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
2000 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2001 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2002 uint8_t halfHV[64];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2003 copy_block9(full, src, 16, stride, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2004 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2005 put ## RND ## pixels8_l2(halfH, halfH, full, 8, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2006 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2007 OPNAME ## pixels8_l2(dst, halfH+8, halfHV, stride, 8, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2008 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2009 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
2010 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2011 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2012 uint8_t halfV[64];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2013 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2014 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2015 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full , 8, 16, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2016 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full+1, 8, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2017 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2018 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
2019 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2020 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
2021 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2022 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2023 uint8_t halfHV[64];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2024 copy_block9(full, src, 16, stride, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2025 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2026 put ## RND ## pixels8_l2(halfH, halfH, full+1, 8, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2027 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2028 OPNAME ## pixels8_l2(dst, halfH+8, halfHV, stride, 8, 8, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2029 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2030 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
2031 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2032 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2033 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, src, 8, stride, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2034 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2035 OPNAME ## pixels8_l2(dst, halfH, halfHV, stride, 8, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2036 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2037 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
2038 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2039 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2040 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, src, 8, stride, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2041 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2042 OPNAME ## pixels8_l2(dst, halfH+8, halfHV, stride, 8, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2043 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2044 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
2045 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2046 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2047 uint8_t halfV[64];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2048 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2049 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2050 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2051 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full, 8, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2052 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2053 OPNAME ## pixels8_l2(dst, halfV, halfHV, stride, 8, 8, 8);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
2054 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2055 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
2056 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2057 uint8_t halfH[72];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2058 copy_block9(full, src, 16, stride, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2059 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2060 put ## RND ## pixels8_l2(halfH, halfH, full, 8, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2061 OPNAME ## mpeg4_qpel8_v_lowpass(dst, halfH, stride, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2062 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2063 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
2064 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2065 uint8_t halfH[72];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2066 uint8_t halfV[64];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2067 uint8_t halfHV[64];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2068 copy_block9(full, src, 16, stride, 9);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2069 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2070 put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full+1, 8, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2071 put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2072 OPNAME ## pixels8_l2(dst, halfV, halfHV, stride, 8, 8, 8);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2073 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2074 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
2075 uint8_t full[16*9];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2076 uint8_t halfH[72];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2077 copy_block9(full, src, 16, stride, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2078 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2079 put ## RND ## pixels8_l2(halfH, halfH, full+1, 8, 8, 16, 9);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2080 OPNAME ## mpeg4_qpel8_v_lowpass(dst, halfH, stride, 8);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2081 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2082 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
2083 uint8_t halfH[72];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2084 put ## RND ## mpeg4_qpel8_h_lowpass(halfH, src, 8, stride, 9);\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2085 OPNAME ## mpeg4_qpel8_v_lowpass(dst, halfH, stride, 8);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2086 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2087 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
2088 OPNAME ## pixels16_c(dst, src, stride, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
2089 }\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2090 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2091 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
2092 uint8_t half[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2093 put ## RND ## mpeg4_qpel16_h_lowpass(half, src, 16, stride, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2094 OPNAME ## pixels16_l2(dst, src, half, stride, stride, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2095 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2096 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2097 static void OPNAME ## qpel16_mc20_c(uint8_t *dst, uint8_t *src, int stride){\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2098 OPNAME ## mpeg4_qpel16_h_lowpass(dst, src, stride, stride, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2099 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2100 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2101 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
2102 uint8_t half[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2103 put ## RND ## mpeg4_qpel16_h_lowpass(half, src, 16, stride, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2104 OPNAME ## pixels16_l2(dst, src+1, half, stride, stride, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2105 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2106 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2107 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
2108 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2109 uint8_t half[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2110 copy_block17(full, src, 24, stride, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2111 put ## RND ## mpeg4_qpel16_v_lowpass(half, full, 16, 24);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2112 OPNAME ## pixels16_l2(dst, full, half, stride, 24, 16, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
2113 }\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2114 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2115 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
2116 uint8_t full[24*17];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2117 copy_block17(full, src, 24, stride, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2118 OPNAME ## mpeg4_qpel16_v_lowpass(dst, full, stride, 24);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2119 }\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2120 \
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2121 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
2122 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2123 uint8_t half[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2124 copy_block17(full, src, 24, stride, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2125 put ## RND ## mpeg4_qpel16_v_lowpass(half, full, 16, 24);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2126 OPNAME ## pixels16_l2(dst, full+24, half, stride, 24, 16, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
2127 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2128 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
2129 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2130 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2131 uint8_t halfV[256];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2132 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2133 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2134 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2135 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full, 16, 24);\
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2136 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2137 OPNAME ## pixels16_l4(dst, full, halfH, halfV, halfHV, stride, 24, 16, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2138 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2139 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
2140 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2141 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2142 uint8_t halfHV[256];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2143 copy_block17(full, src, 24, stride, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2144 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2145 put ## RND ## pixels16_l2(halfH, halfH, full, 16, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2146 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2147 OPNAME ## pixels16_l2(dst, halfH, halfHV, stride, 16, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2148 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2149 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
2150 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2151 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2152 uint8_t halfV[256];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2153 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2154 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2155 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2156 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full+1, 16, 24);\
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2157 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2158 OPNAME ## pixels16_l4(dst, full+1, halfH, halfV, halfHV, stride, 24, 16, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2159 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2160 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
2161 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2162 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2163 uint8_t halfHV[256];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2164 copy_block17(full, src, 24, stride, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2165 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2166 put ## RND ## pixels16_l2(halfH, halfH, full+1, 16, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2167 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2168 OPNAME ## pixels16_l2(dst, halfH, halfHV, stride, 16, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2169 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2170 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
2171 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2172 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2173 uint8_t halfV[256];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2174 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2175 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2176 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2177 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full, 16, 24);\
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2178 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2179 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
2180 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2181 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
2182 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2183 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2184 uint8_t halfHV[256];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2185 copy_block17(full, src, 24, stride, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2186 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2187 put ## RND ## pixels16_l2(halfH, halfH, full, 16, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2188 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2189 OPNAME ## pixels16_l2(dst, halfH+16, halfHV, stride, 16, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2190 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2191 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
2192 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2193 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2194 uint8_t halfV[256];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2195 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2196 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2197 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full , 16, 24, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2198 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full+1, 16, 24);\
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2199 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2200 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
2201 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2202 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
2203 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2204 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2205 uint8_t halfHV[256];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2206 copy_block17(full, src, 24, stride, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2207 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2208 put ## RND ## pixels16_l2(halfH, halfH, full+1, 16, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2209 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2210 OPNAME ## pixels16_l2(dst, halfH+16, halfHV, stride, 16, 16, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2211 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2212 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
2213 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2214 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2215 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, src, 16, stride, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2216 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2217 OPNAME ## pixels16_l2(dst, halfH, halfHV, stride, 16, 16, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
2218 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2219 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
2220 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2221 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2222 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, src, 16, stride, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2223 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2224 OPNAME ## pixels16_l2(dst, halfH+16, halfHV, stride, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2225 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2226 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
2227 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2228 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2229 uint8_t halfV[256];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2230 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2231 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2232 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2233 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full, 16, 24);\
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2234 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2235 OPNAME ## pixels16_l2(dst, halfV, halfHV, stride, 16, 16, 16);\
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
2236 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2237 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
2238 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2239 uint8_t halfH[272];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2240 copy_block17(full, src, 24, stride, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2241 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2242 put ## RND ## pixels16_l2(halfH, halfH, full, 16, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2243 OPNAME ## mpeg4_qpel16_v_lowpass(dst, halfH, stride, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2244 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2245 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
2246 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2247 uint8_t halfH[272];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2248 uint8_t halfV[256];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2249 uint8_t halfHV[256];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2250 copy_block17(full, src, 24, stride, 17);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2251 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2252 put ## RND ## mpeg4_qpel16_v_lowpass(halfV, full+1, 16, 24);\
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2253 put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2254 OPNAME ## pixels16_l2(dst, halfV, halfHV, stride, 16, 16, 16);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2255 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2256 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
2257 uint8_t full[24*17];\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2258 uint8_t halfH[272];\
984
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2259 copy_block17(full, src, 24, stride, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2260 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2261 put ## RND ## pixels16_l2(halfH, halfH, full+1, 16, 16, 24, 17);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2262 OPNAME ## mpeg4_qpel16_v_lowpass(dst, halfH, stride, 16);\
e162c09efbe7 qpel fix
michaelni
parents: 954
diff changeset
2263 }\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
2264 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
2265 uint8_t halfH[272];\
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2266 put ## RND ## mpeg4_qpel16_h_lowpass(halfH, src, 16, stride, 17);\
954
13aec7e50c52 qpel in mmx2/3dnow
michaelni
parents: 945
diff changeset
2267 OPNAME ## mpeg4_qpel16_v_lowpass(dst, halfH, stride, 16);\
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
2268 }
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
2269
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2270 #define op_avg(a, b) a = (((a)+cm[((b) + 16)>>5]+1)>>1)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2271 #define op_avg_no_rnd(a, b) a = (((a)+cm[((b) + 15)>>5])>>1)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2272 #define op_put(a, b) a = cm[((b) + 16)>>5]
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2273 #define op_put_no_rnd(a, b) a = cm[((b) + 15)>>5]
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2274
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2275 QPEL_MC(0, put_ , _ , op_put)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2276 QPEL_MC(1, put_no_rnd_, _no_rnd_, op_put_no_rnd)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2277 QPEL_MC(0, avg_ , _ , op_avg)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2278 //QPEL_MC(1, avg_no_rnd , _ , op_avg)
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2279 #undef op_avg
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2280 #undef op_avg_no_rnd
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2281 #undef op_put
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
2282 #undef op_put_no_rnd
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
2283
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2284 #if 1
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2285 #define H264_LOWPASS(OPNAME, OP, OP2) \
5151
8c273ff9ca03 kill some "defined but not used" warnings
mru
parents: 5150
diff changeset
2286 static av_unused void OPNAME ## h264_qpel2_h_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
3020
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2287 const int h=2;\
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
2288 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;\
3020
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2289 int i;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2290 for(i=0; i<h; i++)\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2291 {\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2292 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
2293 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
2294 dst+=dstStride;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2295 src+=srcStride;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2296 }\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2297 }\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2298 \
5151
8c273ff9ca03 kill some "defined but not used" warnings
mru
parents: 5150
diff changeset
2299 static av_unused void OPNAME ## h264_qpel2_v_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
3020
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2300 const int w=2;\
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
2301 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;\
3020
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2302 int i;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2303 for(i=0; i<w; i++)\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2304 {\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2305 const int srcB= src[-2*srcStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2306 const int srcA= src[-1*srcStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2307 const int src0= src[0 *srcStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2308 const int src1= src[1 *srcStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2309 const int src2= src[2 *srcStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2310 const int src3= src[3 *srcStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2311 const int src4= src[4 *srcStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2312 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
2313 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
2314 dst++;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2315 src++;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2316 }\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2317 }\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2318 \
5151
8c273ff9ca03 kill some "defined but not used" warnings
mru
parents: 5150
diff changeset
2319 static av_unused void OPNAME ## h264_qpel2_hv_lowpass(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
3020
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2320 const int h=2;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2321 const int w=2;\
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
2322 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;\
3020
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2323 int i;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2324 src -= 2*srcStride;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2325 for(i=0; i<h+5; i++)\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2326 {\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2327 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
2328 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
2329 tmp+=tmpStride;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2330 src+=srcStride;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2331 }\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2332 tmp -= tmpStride*(h+5-2);\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2333 for(i=0; i<w; i++)\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2334 {\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2335 const int tmpB= tmp[-2*tmpStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2336 const int tmpA= tmp[-1*tmpStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2337 const int tmp0= tmp[0 *tmpStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2338 const int tmp1= tmp[1 *tmpStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2339 const int tmp2= tmp[2 *tmpStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2340 const int tmp3= tmp[3 *tmpStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2341 const int tmp4= tmp[4 *tmpStride];\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2342 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
2343 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
2344 dst++;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2345 tmp++;\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2346 }\
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
2347 }\
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2348 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
2349 const int h=4;\
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
2350 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;\
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2351 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2352 for(i=0; i<h; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2353 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2354 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
2355 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
2356 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
2357 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
2358 dst+=dstStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2359 src+=srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2360 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2361 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2362 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2363 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
2364 const int w=4;\
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
2365 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;\
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2366 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2367 for(i=0; i<w; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2368 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2369 const int srcB= src[-2*srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2370 const int srcA= src[-1*srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2371 const int src0= src[0 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2372 const int src1= src[1 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2373 const int src2= src[2 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2374 const int src3= src[3 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2375 const int src4= src[4 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2376 const int src5= src[5 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2377 const int src6= src[6 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2378 OP(dst[0*dstStride], (src0+src1)*20 - (srcA+src2)*5 + (srcB+src3));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2379 OP(dst[1*dstStride], (src1+src2)*20 - (src0+src3)*5 + (srcA+src4));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2380 OP(dst[2*dstStride], (src2+src3)*20 - (src1+src4)*5 + (src0+src5));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2381 OP(dst[3*dstStride], (src3+src4)*20 - (src2+src5)*5 + (src1+src6));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2382 dst++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2383 src++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2384 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2385 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2386 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2387 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
2388 const int h=4;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2389 const int w=4;\
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
2390 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;\
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2391 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2392 src -= 2*srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2393 for(i=0; i<h+5; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2394 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2395 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
2396 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
2397 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
2398 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
2399 tmp+=tmpStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2400 src+=srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2401 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2402 tmp -= tmpStride*(h+5-2);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2403 for(i=0; i<w; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2404 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2405 const int tmpB= tmp[-2*tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2406 const int tmpA= tmp[-1*tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2407 const int tmp0= tmp[0 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2408 const int tmp1= tmp[1 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2409 const int tmp2= tmp[2 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2410 const int tmp3= tmp[3 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2411 const int tmp4= tmp[4 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2412 const int tmp5= tmp[5 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2413 const int tmp6= tmp[6 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2414 OP2(dst[0*dstStride], (tmp0+tmp1)*20 - (tmpA+tmp2)*5 + (tmpB+tmp3));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2415 OP2(dst[1*dstStride], (tmp1+tmp2)*20 - (tmp0+tmp3)*5 + (tmpA+tmp4));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2416 OP2(dst[2*dstStride], (tmp2+tmp3)*20 - (tmp1+tmp4)*5 + (tmp0+tmp5));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2417 OP2(dst[3*dstStride], (tmp3+tmp4)*20 - (tmp2+tmp5)*5 + (tmp1+tmp6));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2418 dst++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2419 tmp++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2420 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2421 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2422 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2423 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
2424 const int h=8;\
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
2425 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;\
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2426 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2427 for(i=0; i<h; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2428 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2429 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
2430 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
2431 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
2432 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
2433 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
2434 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
2435 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
2436 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
2437 dst+=dstStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2438 src+=srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2439 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2440 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2441 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2442 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
2443 const int w=8;\
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
2444 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;\
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2445 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2446 for(i=0; i<w; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2447 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2448 const int srcB= src[-2*srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2449 const int srcA= src[-1*srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2450 const int src0= src[0 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2451 const int src1= src[1 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2452 const int src2= src[2 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2453 const int src3= src[3 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2454 const int src4= src[4 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2455 const int src5= src[5 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2456 const int src6= src[6 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2457 const int src7= src[7 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2458 const int src8= src[8 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2459 const int src9= src[9 *srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2460 const int src10=src[10*srcStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2461 OP(dst[0*dstStride], (src0+src1)*20 - (srcA+src2)*5 + (srcB+src3));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2462 OP(dst[1*dstStride], (src1+src2)*20 - (src0+src3)*5 + (srcA+src4));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2463 OP(dst[2*dstStride], (src2+src3)*20 - (src1+src4)*5 + (src0+src5));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2464 OP(dst[3*dstStride], (src3+src4)*20 - (src2+src5)*5 + (src1+src6));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2465 OP(dst[4*dstStride], (src4+src5)*20 - (src3+src6)*5 + (src2+src7));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2466 OP(dst[5*dstStride], (src5+src6)*20 - (src4+src7)*5 + (src3+src8));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2467 OP(dst[6*dstStride], (src6+src7)*20 - (src5+src8)*5 + (src4+src9));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2468 OP(dst[7*dstStride], (src7+src8)*20 - (src6+src9)*5 + (src5+src10));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2469 dst++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2470 src++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2471 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2472 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2473 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2474 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
2475 const int h=8;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2476 const int w=8;\
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
2477 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;\
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2478 int i;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2479 src -= 2*srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2480 for(i=0; i<h+5; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2481 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2482 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
2483 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
2484 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
2485 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
2486 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
2487 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
2488 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
2489 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
2490 tmp+=tmpStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2491 src+=srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2492 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2493 tmp -= tmpStride*(h+5-2);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2494 for(i=0; i<w; i++)\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2495 {\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2496 const int tmpB= tmp[-2*tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2497 const int tmpA= tmp[-1*tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2498 const int tmp0= tmp[0 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2499 const int tmp1= tmp[1 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2500 const int tmp2= tmp[2 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2501 const int tmp3= tmp[3 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2502 const int tmp4= tmp[4 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2503 const int tmp5= tmp[5 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2504 const int tmp6= tmp[6 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2505 const int tmp7= tmp[7 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2506 const int tmp8= tmp[8 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2507 const int tmp9= tmp[9 *tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2508 const int tmp10=tmp[10*tmpStride];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2509 OP2(dst[0*dstStride], (tmp0+tmp1)*20 - (tmpA+tmp2)*5 + (tmpB+tmp3));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2510 OP2(dst[1*dstStride], (tmp1+tmp2)*20 - (tmp0+tmp3)*5 + (tmpA+tmp4));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2511 OP2(dst[2*dstStride], (tmp2+tmp3)*20 - (tmp1+tmp4)*5 + (tmp0+tmp5));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2512 OP2(dst[3*dstStride], (tmp3+tmp4)*20 - (tmp2+tmp5)*5 + (tmp1+tmp6));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2513 OP2(dst[4*dstStride], (tmp4+tmp5)*20 - (tmp3+tmp6)*5 + (tmp2+tmp7));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2514 OP2(dst[5*dstStride], (tmp5+tmp6)*20 - (tmp4+tmp7)*5 + (tmp3+tmp8));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2515 OP2(dst[6*dstStride], (tmp6+tmp7)*20 - (tmp5+tmp8)*5 + (tmp4+tmp9));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2516 OP2(dst[7*dstStride], (tmp7+tmp8)*20 - (tmp6+tmp9)*5 + (tmp5+tmp10));\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2517 dst++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2518 tmp++;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2519 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2520 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2521 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2522 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
2523 OPNAME ## h264_qpel8_v_lowpass(dst , src , dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2524 OPNAME ## h264_qpel8_v_lowpass(dst+8, src+8, dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2525 src += 8*srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2526 dst += 8*dstStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2527 OPNAME ## h264_qpel8_v_lowpass(dst , src , dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2528 OPNAME ## h264_qpel8_v_lowpass(dst+8, src+8, dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2529 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2530 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2531 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
2532 OPNAME ## h264_qpel8_h_lowpass(dst , src , dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2533 OPNAME ## h264_qpel8_h_lowpass(dst+8, src+8, dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2534 src += 8*srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2535 dst += 8*dstStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2536 OPNAME ## h264_qpel8_h_lowpass(dst , src , dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2537 OPNAME ## h264_qpel8_h_lowpass(dst+8, src+8, dstStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2538 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2539 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2540 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
2541 OPNAME ## h264_qpel8_hv_lowpass(dst , tmp , src , dstStride, tmpStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2542 OPNAME ## h264_qpel8_hv_lowpass(dst+8, tmp+8, src+8, dstStride, tmpStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2543 src += 8*srcStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2544 dst += 8*dstStride;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2545 OPNAME ## h264_qpel8_hv_lowpass(dst , tmp , src , dstStride, tmpStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2546 OPNAME ## h264_qpel8_hv_lowpass(dst+8, tmp+8, src+8, dstStride, tmpStride, srcStride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2547 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2548
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2549 #define H264_MC(OPNAME, SIZE) \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2550 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
2551 OPNAME ## pixels ## SIZE ## _c(dst, src, stride, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2552 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2553 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2554 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
2555 uint8_t half[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2556 put_h264_qpel ## SIZE ## _h_lowpass(half, src, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2557 OPNAME ## pixels ## SIZE ## _l2(dst, src, half, stride, stride, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2558 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2559 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2560 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
2561 OPNAME ## h264_qpel ## SIZE ## _h_lowpass(dst, src, stride, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2562 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2563 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2564 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
2565 uint8_t half[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2566 put_h264_qpel ## SIZE ## _h_lowpass(half, src, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2567 OPNAME ## pixels ## SIZE ## _l2(dst, src+1, half, stride, stride, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2568 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2569 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2570 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
2571 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2572 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2573 uint8_t half[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2574 copy_block ## SIZE (full, src - stride*2, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2575 put_h264_qpel ## SIZE ## _v_lowpass(half, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2576 OPNAME ## pixels ## SIZE ## _l2(dst, full_mid, half, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2577 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2578 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2579 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
2580 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2581 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2582 copy_block ## SIZE (full, src - stride*2, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2583 OPNAME ## h264_qpel ## SIZE ## _v_lowpass(dst, full_mid, stride, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2584 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2585 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2586 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
2587 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2588 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2589 uint8_t half[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2590 copy_block ## SIZE (full, src - stride*2, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2591 put_h264_qpel ## SIZE ## _v_lowpass(half, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2592 OPNAME ## pixels ## SIZE ## _l2(dst, full_mid+SIZE, half, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2593 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2594 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2595 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
2596 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2597 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2598 uint8_t halfH[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2599 uint8_t halfV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2600 put_h264_qpel ## SIZE ## _h_lowpass(halfH, src, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2601 copy_block ## SIZE (full, src - stride*2, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2602 put_h264_qpel ## SIZE ## _v_lowpass(halfV, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2603 OPNAME ## pixels ## SIZE ## _l2(dst, halfH, halfV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2604 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2605 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2606 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
2607 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2608 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2609 uint8_t halfH[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2610 uint8_t halfV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2611 put_h264_qpel ## SIZE ## _h_lowpass(halfH, src, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2612 copy_block ## SIZE (full, src - stride*2 + 1, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2613 put_h264_qpel ## SIZE ## _v_lowpass(halfV, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2614 OPNAME ## pixels ## SIZE ## _l2(dst, halfH, halfV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2615 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2616 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2617 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
2618 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2619 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2620 uint8_t halfH[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2621 uint8_t halfV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2622 put_h264_qpel ## SIZE ## _h_lowpass(halfH, src + stride, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2623 copy_block ## SIZE (full, src - stride*2, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2624 put_h264_qpel ## SIZE ## _v_lowpass(halfV, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2625 OPNAME ## pixels ## SIZE ## _l2(dst, halfH, halfV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2626 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2627 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2628 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
2629 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2630 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2631 uint8_t halfH[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2632 uint8_t halfV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2633 put_h264_qpel ## SIZE ## _h_lowpass(halfH, src + stride, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2634 copy_block ## SIZE (full, src - stride*2 + 1, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2635 put_h264_qpel ## SIZE ## _v_lowpass(halfV, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2636 OPNAME ## pixels ## SIZE ## _l2(dst, halfH, halfV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2637 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2638 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2639 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
2640 int16_t tmp[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2641 OPNAME ## h264_qpel ## SIZE ## _hv_lowpass(dst, tmp, src, stride, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2642 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2643 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2644 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
2645 int16_t tmp[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2646 uint8_t halfH[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2647 uint8_t halfHV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2648 put_h264_qpel ## SIZE ## _h_lowpass(halfH, src, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2649 put_h264_qpel ## SIZE ## _hv_lowpass(halfHV, tmp, src, SIZE, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2650 OPNAME ## pixels ## SIZE ## _l2(dst, halfH, halfHV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2651 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2652 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2653 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
2654 int16_t tmp[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2655 uint8_t halfH[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2656 uint8_t halfHV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2657 put_h264_qpel ## SIZE ## _h_lowpass(halfH, src + stride, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2658 put_h264_qpel ## SIZE ## _hv_lowpass(halfHV, tmp, src, SIZE, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2659 OPNAME ## pixels ## SIZE ## _l2(dst, halfH, halfHV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2660 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2661 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2662 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
2663 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2664 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2665 int16_t tmp[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2666 uint8_t halfV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2667 uint8_t halfHV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2668 copy_block ## SIZE (full, src - stride*2, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2669 put_h264_qpel ## SIZE ## _v_lowpass(halfV, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2670 put_h264_qpel ## SIZE ## _hv_lowpass(halfHV, tmp, src, SIZE, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2671 OPNAME ## pixels ## SIZE ## _l2(dst, halfV, halfHV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2672 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2673 \
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2674 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
2675 uint8_t full[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2676 uint8_t * const full_mid= full + SIZE*2;\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2677 int16_t tmp[SIZE*(SIZE+5)];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2678 uint8_t halfV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2679 uint8_t halfHV[SIZE*SIZE];\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2680 copy_block ## SIZE (full, src - stride*2 + 1, SIZE, stride, SIZE + 5);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2681 put_h264_qpel ## SIZE ## _v_lowpass(halfV, full_mid, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2682 put_h264_qpel ## SIZE ## _hv_lowpass(halfHV, tmp, src, SIZE, SIZE, stride);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2683 OPNAME ## pixels ## SIZE ## _l2(dst, halfV, halfHV, stride, SIZE, SIZE, SIZE);\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2684 }\
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2685
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2686 #define op_avg(a, b) a = (((a)+cm[((b) + 16)>>5]+1)>>1)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2687 //#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
2688 #define op_put(a, b) a = cm[((b) + 16)>>5]
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2689 #define op2_avg(a, b) a = (((a)+cm[((b) + 512)>>10]+1)>>1)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2690 #define op2_put(a, b) a = cm[((b) + 512)>>10]
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2691
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2692 H264_LOWPASS(put_ , op_put, op2_put)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2693 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
2694 H264_MC(put_, 2)
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2695 H264_MC(put_, 4)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2696 H264_MC(put_, 8)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2697 H264_MC(put_, 16)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2698 H264_MC(avg_, 4)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2699 H264_MC(avg_, 8)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2700 H264_MC(avg_, 16)
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2701
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2702 #undef op_avg
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2703 #undef op_put
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2704 #undef op2_avg
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2705 #undef op2_put
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2706 #endif
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
2707
4594
a96d905dcbaa Add av_ prefix to clip functions
reimar
parents: 4296
diff changeset
2708 #define op_scale1(x) block[x] = av_clip_uint8( (block[x]*weight + offset) >> log2_denom )
a96d905dcbaa Add av_ prefix to clip functions
reimar
parents: 4296
diff changeset
2709 #define op_scale2(x) dst[x] = av_clip_uint8( (src[x]*weights + dst[x]*weightd + offset) >> (log2_denom+1))
2415
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2710 #define H264_WEIGHT(W,H) \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2711 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
2712 int y; \
2415
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2713 offset <<= log2_denom; \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2714 if(log2_denom) offset += 1<<(log2_denom-1); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2715 for(y=0; y<H; y++, block += stride){ \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2716 op_scale1(0); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2717 op_scale1(1); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2718 if(W==2) continue; \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2719 op_scale1(2); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2720 op_scale1(3); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2721 if(W==4) continue; \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2722 op_scale1(4); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2723 op_scale1(5); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2724 op_scale1(6); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2725 op_scale1(7); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2726 if(W==8) continue; \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2727 op_scale1(8); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2728 op_scale1(9); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2729 op_scale1(10); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2730 op_scale1(11); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2731 op_scale1(12); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2732 op_scale1(13); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2733 op_scale1(14); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2734 op_scale1(15); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2735 } \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2736 } \
3029
7a92269a3ccd tweak h264_biweight
lorenm
parents: 3020
diff changeset
2737 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
2738 int y; \
7a92269a3ccd tweak h264_biweight
lorenm
parents: 3020
diff changeset
2739 offset = ((offset + 1) | 1) << log2_denom; \
2415
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2740 for(y=0; y<H; y++, dst += stride, src += stride){ \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2741 op_scale2(0); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2742 op_scale2(1); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2743 if(W==2) continue; \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2744 op_scale2(2); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2745 op_scale2(3); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2746 if(W==4) continue; \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2747 op_scale2(4); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2748 op_scale2(5); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2749 op_scale2(6); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2750 op_scale2(7); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2751 if(W==8) continue; \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2752 op_scale2(8); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2753 op_scale2(9); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2754 op_scale2(10); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2755 op_scale2(11); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2756 op_scale2(12); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2757 op_scale2(13); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2758 op_scale2(14); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2759 op_scale2(15); \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2760 } \
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2761 }
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2762
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2763 H264_WEIGHT(16,16)
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2764 H264_WEIGHT(16,8)
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2765 H264_WEIGHT(8,16)
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2766 H264_WEIGHT(8,8)
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2767 H264_WEIGHT(8,4)
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2768 H264_WEIGHT(4,8)
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2769 H264_WEIGHT(4,4)
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2770 H264_WEIGHT(4,2)
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2771 H264_WEIGHT(2,4)
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2772 H264_WEIGHT(2,2)
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2773
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2774 #undef op_scale1
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2775 #undef op_scale2
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2776 #undef H264_WEIGHT
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
2777
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2778 static void wmv2_mspel8_h_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
2779 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2780 int i;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2781
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2782 for(i=0; i<h; i++){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2783 dst[0]= cm[(9*(src[0] + src[1]) - (src[-1] + src[2]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2784 dst[1]= cm[(9*(src[1] + src[2]) - (src[ 0] + src[3]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2785 dst[2]= cm[(9*(src[2] + src[3]) - (src[ 1] + src[4]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2786 dst[3]= cm[(9*(src[3] + src[4]) - (src[ 2] + src[5]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2787 dst[4]= cm[(9*(src[4] + src[5]) - (src[ 3] + src[6]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2788 dst[5]= cm[(9*(src[5] + src[6]) - (src[ 4] + src[7]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2789 dst[6]= cm[(9*(src[6] + src[7]) - (src[ 5] + src[8]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2790 dst[7]= cm[(9*(src[7] + src[8]) - (src[ 6] + src[9]) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2791 dst+=dstStride;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2792 src+=srcStride;
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2793 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2794 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2795
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8496
diff changeset
2796 #if CONFIG_CAVS_DECODER
3395
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2797 /* AVS specific */
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2798 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
2799 put_pixels8_c(dst, src, stride, 8);
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2800 }
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2801 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
2802 avg_pixels8_c(dst, src, stride, 8);
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2803 }
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2804 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
2805 put_pixels16_c(dst, src, stride, 16);
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2806 }
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2807 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
2808 avg_pixels16_c(dst, src, stride, 16);
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2809 }
3432
ab49baf4adad Fix compilation with --disable-decoders.
diego
parents: 3395
diff changeset
2810 #endif /* CONFIG_CAVS_DECODER */
3395
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
2811
9995
3141f69e3905 Do not check for both CONFIG_VC1_DECODER and CONFIG_WMV3_DECODER,
diego
parents: 9975
diff changeset
2812 #if CONFIG_VC1_DECODER
3526
7dc8e4a12105 New functions in DSPContext for VC-1 decoding
kostya
parents: 3432
diff changeset
2813 /* VC-1 specific */
7dc8e4a12105 New functions in DSPContext for VC-1 decoding
kostya
parents: 3432
diff changeset
2814 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
2815 put_pixels8_c(dst, src, stride, 8);
7dc8e4a12105 New functions in DSPContext for VC-1 decoding
kostya
parents: 3432
diff changeset
2816 }
9437
8aa2e86549cd VC1: Do qpel when needed for both MVs in a B frame
conrad
parents: 9259
diff changeset
2817 void ff_avg_vc1_mspel_mc00_c(uint8_t *dst, uint8_t *src, int stride, int rnd) {
8aa2e86549cd VC1: Do qpel when needed for both MVs in a B frame
conrad
parents: 9259
diff changeset
2818 avg_pixels8_c(dst, src, stride, 8);
8aa2e86549cd VC1: Do qpel when needed for both MVs in a B frame
conrad
parents: 9259
diff changeset
2819 }
9995
3141f69e3905 Do not check for both CONFIG_VC1_DECODER and CONFIG_WMV3_DECODER,
diego
parents: 9975
diff changeset
2820 #endif /* CONFIG_VC1_DECODER */
3526
7dc8e4a12105 New functions in DSPContext for VC-1 decoding
kostya
parents: 3432
diff changeset
2821
4296
4bfaef73d6ed Adding H.264 DCT.
takis
parents: 4292
diff changeset
2822 /* H264 specific */
5411
362aec4ef932 Take care of some renames (Doxygen and function name) after the previous pure rename patch.
takis
parents: 5394
diff changeset
2823 void ff_h264dspenc_init(DSPContext* c, AVCodecContext *avctx);
4296
4bfaef73d6ed Adding H.264 DCT.
takis
parents: 4292
diff changeset
2824
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8496
diff changeset
2825 #if CONFIG_RV40_DECODER
8232
520a6e56b8c0 RV40 MC functions
kostya
parents: 8120
diff changeset
2826 static void put_rv40_qpel16_mc33_c(uint8_t *dst, uint8_t *src, int stride){
520a6e56b8c0 RV40 MC functions
kostya
parents: 8120
diff changeset
2827 put_pixels16_xy2_c(dst, src, stride, 16);
520a6e56b8c0 RV40 MC functions
kostya
parents: 8120
diff changeset
2828 }
520a6e56b8c0 RV40 MC functions
kostya
parents: 8120
diff changeset
2829 static void avg_rv40_qpel16_mc33_c(uint8_t *dst, uint8_t *src, int stride){
520a6e56b8c0 RV40 MC functions
kostya
parents: 8120
diff changeset
2830 avg_pixels16_xy2_c(dst, src, stride, 16);
520a6e56b8c0 RV40 MC functions
kostya
parents: 8120
diff changeset
2831 }
520a6e56b8c0 RV40 MC functions
kostya
parents: 8120
diff changeset
2832 static void put_rv40_qpel8_mc33_c(uint8_t *dst, uint8_t *src, int stride){
520a6e56b8c0 RV40 MC functions
kostya
parents: 8120
diff changeset
2833 put_pixels8_xy2_c(dst, src, stride, 8);
520a6e56b8c0 RV40 MC functions
kostya
parents: 8120
diff changeset
2834 }
520a6e56b8c0 RV40 MC functions
kostya
parents: 8120
diff changeset
2835 static void avg_rv40_qpel8_mc33_c(uint8_t *dst, uint8_t *src, int stride){
520a6e56b8c0 RV40 MC functions
kostya
parents: 8120
diff changeset
2836 avg_pixels8_xy2_c(dst, src, stride, 8);
520a6e56b8c0 RV40 MC functions
kostya
parents: 8120
diff changeset
2837 }
520a6e56b8c0 RV40 MC functions
kostya
parents: 8120
diff changeset
2838 #endif /* CONFIG_RV40_DECODER */
520a6e56b8c0 RV40 MC functions
kostya
parents: 8120
diff changeset
2839
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2840 static void wmv2_mspel8_v_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int w){
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
2841 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2842 int i;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2843
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2844 for(i=0; i<w; i++){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2845 const int src_1= src[ -srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2846 const int src0 = src[0 ];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2847 const int src1 = src[ srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2848 const int src2 = src[2*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2849 const int src3 = src[3*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2850 const int src4 = src[4*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2851 const int src5 = src[5*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2852 const int src6 = src[6*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2853 const int src7 = src[7*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2854 const int src8 = src[8*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2855 const int src9 = src[9*srcStride];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2856 dst[0*dstStride]= cm[(9*(src0 + src1) - (src_1 + src2) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2857 dst[1*dstStride]= cm[(9*(src1 + src2) - (src0 + src3) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2858 dst[2*dstStride]= cm[(9*(src2 + src3) - (src1 + src4) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2859 dst[3*dstStride]= cm[(9*(src3 + src4) - (src2 + src5) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2860 dst[4*dstStride]= cm[(9*(src4 + src5) - (src3 + src6) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2861 dst[5*dstStride]= cm[(9*(src5 + src6) - (src4 + src7) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2862 dst[6*dstStride]= cm[(9*(src6 + src7) - (src5 + src8) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2863 dst[7*dstStride]= cm[(9*(src7 + src8) - (src6 + src9) + 8)>>4];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2864 src++;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2865 dst++;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2866 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2867 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2868
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2869 static void put_mspel8_mc00_c (uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2870 put_pixels8_c(dst, src, stride, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2871 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2872
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2873 static void put_mspel8_mc10_c(uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2874 uint8_t half[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2875 wmv2_mspel8_h_lowpass(half, src, 8, stride, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2876 put_pixels8_l2(dst, src, half, stride, stride, 8, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2877 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2878
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2879 static void put_mspel8_mc20_c(uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2880 wmv2_mspel8_h_lowpass(dst, src, stride, stride, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2881 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2882
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2883 static void put_mspel8_mc30_c(uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2884 uint8_t half[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2885 wmv2_mspel8_h_lowpass(half, src, 8, stride, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2886 put_pixels8_l2(dst, src+1, half, stride, stride, 8, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2887 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2888
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2889 static void put_mspel8_mc02_c(uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2890 wmv2_mspel8_v_lowpass(dst, src, stride, stride, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2891 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2892
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2893 static void put_mspel8_mc12_c(uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2894 uint8_t halfH[88];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2895 uint8_t halfV[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2896 uint8_t halfHV[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2897 wmv2_mspel8_h_lowpass(halfH, src-stride, 8, stride, 11);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2898 wmv2_mspel8_v_lowpass(halfV, src, 8, stride, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2899 wmv2_mspel8_v_lowpass(halfHV, halfH+8, 8, 8, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2900 put_pixels8_l2(dst, halfV, halfHV, stride, 8, 8, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2901 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2902 static void put_mspel8_mc32_c(uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2903 uint8_t halfH[88];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2904 uint8_t halfV[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2905 uint8_t halfHV[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2906 wmv2_mspel8_h_lowpass(halfH, src-stride, 8, stride, 11);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2907 wmv2_mspel8_v_lowpass(halfV, src+1, 8, stride, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2908 wmv2_mspel8_v_lowpass(halfHV, halfH+8, 8, 8, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2909 put_pixels8_l2(dst, halfV, halfHV, stride, 8, 8, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2910 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2911 static void put_mspel8_mc22_c(uint8_t *dst, uint8_t *src, int stride){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2912 uint8_t halfH[88];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2913 wmv2_mspel8_h_lowpass(halfH, src-stride, 8, stride, 11);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2914 wmv2_mspel8_v_lowpass(dst, halfH+8, stride, 8, 8);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2915 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2916
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2917 static void h263_v_loop_filter_c(uint8_t *src, int stride, int qscale){
10749
5cca4b6c459d Get rid of pointless CONFIG_ANY_H263 preprocessor definition.
diego
parents: 10748
diff changeset
2918 if(CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2919 int x;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2920 const int strength= ff_h263_loop_filter_strength[qscale];
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2921
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2922 for(x=0; x<8; x++){
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2923 int d1, d2, ad1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2924 int p0= src[x-2*stride];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2925 int p1= src[x-1*stride];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2926 int p2= src[x+0*stride];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2927 int p3= src[x+1*stride];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2928 int d = (p0 - p3 + 4*(p2 - p1)) / 8;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2929
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2930 if (d<-2*strength) d1= 0;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2931 else if(d<- strength) d1=-2*strength - d;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2932 else if(d< strength) d1= d;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2933 else if(d< 2*strength) d1= 2*strength - d;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2934 else d1= 0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2935
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2936 p1 += d1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2937 p2 -= d1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2938 if(p1&256) p1= ~(p1>>31);
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2939 if(p2&256) p2= ~(p2>>31);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2940
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2941 src[x-1*stride] = p1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2942 src[x+0*stride] = p2;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2943
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
2944 ad1= FFABS(d1)>>1;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2945
4594
a96d905dcbaa Add av_ prefix to clip functions
reimar
parents: 4296
diff changeset
2946 d2= av_clip((p0-p3)/4, -ad1, ad1);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2947
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2948 src[x-2*stride] = p0 - d2;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2949 src[x+ stride] = p3 + d2;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2950 }
5394
e9a6215f4e3a help some gcc version to optimize out those functions
aurel
parents: 5291
diff changeset
2951 }
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2952 }
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2953
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2954 static void h263_h_loop_filter_c(uint8_t *src, int stride, int qscale){
10749
5cca4b6c459d Get rid of pointless CONFIG_ANY_H263 preprocessor definition.
diego
parents: 10748
diff changeset
2955 if(CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2956 int y;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2957 const int strength= ff_h263_loop_filter_strength[qscale];
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2958
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2959 for(y=0; y<8; y++){
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2960 int d1, d2, ad1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2961 int p0= src[y*stride-2];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2962 int p1= src[y*stride-1];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2963 int p2= src[y*stride+0];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2964 int p3= src[y*stride+1];
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2965 int d = (p0 - p3 + 4*(p2 - p1)) / 8;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2966
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2967 if (d<-2*strength) d1= 0;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2968 else if(d<- strength) d1=-2*strength - d;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2969 else if(d< strength) d1= d;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2970 else if(d< 2*strength) d1= 2*strength - d;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2971 else d1= 0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2972
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2973 p1 += d1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2974 p2 -= d1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2975 if(p1&256) p1= ~(p1>>31);
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2976 if(p2&256) p2= ~(p2>>31);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2977
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2978 src[y*stride-1] = p1;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2979 src[y*stride+0] = p2;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2980
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
2981 ad1= FFABS(d1)>>1;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2982
4594
a96d905dcbaa Add av_ prefix to clip functions
reimar
parents: 4296
diff changeset
2983 d2= av_clip((p0-p3)/4, -ad1, ad1);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
2984
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2985 src[y*stride-2] = p0 - d2;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2986 src[y*stride+1] = p3 + d2;
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2987 }
5394
e9a6215f4e3a help some gcc version to optimize out those functions
aurel
parents: 5291
diff changeset
2988 }
1644
835cf346975e h263 loop filter
michael
parents: 1598
diff changeset
2989 }
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
2990
2045
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2991 static void h261_loop_filter_c(uint8_t *src, int stride){
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2992 int x,y,xy,yz;
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2993 int temp[64];
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2994
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2995 for(x=0; x<8; x++){
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2996 temp[x ] = 4*src[x ];
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2997 temp[x + 7*8] = 4*src[x + 7*stride];
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2998 }
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
2999 for(y=1; y<7; y++){
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
3000 for(x=0; x<8; x++){
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
3001 xy = y * stride + x;
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
3002 yz = y * 8 + x;
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
3003 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
3004 }
b6f2add2511e h261 decoder by (Maarten Daniels <maarten.daniels at student dot luc dot ac dot be>)
michael
parents: 1984
diff changeset
3005 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3006
2045
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
3007 for(y=0; y<8; y++){
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
3008 src[ y*stride] = (temp[ y*8] + 2)>>2;
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
3009 src[7+y*stride] = (temp[7+y*8] + 2)>>2;
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
3010 for(x=1; x<7; x++){
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
3011 xy = y * stride + x;
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
3012 yz = y * 8 + x;
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
3013 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
3014 }
b6f2add2511e h261 decoder by (Maarten Daniels <maarten.daniels at student dot luc dot ac dot be>)
michael
parents: 1984
diff changeset
3015 }
b6f2add2511e h261 decoder by (Maarten Daniels <maarten.daniels at student dot luc dot ac dot be>)
michael
parents: 1984
diff changeset
3016 }
b6f2add2511e h261 decoder by (Maarten Daniels <maarten.daniels at student dot luc dot ac dot be>)
michael
parents: 1984
diff changeset
3017
10941
28edcc8c54c0 Mark the h264 c loop filter as av_always_inline av_flatten to make sure its
michael
parents: 10940
diff changeset
3018 static av_always_inline av_flatten 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
3019 {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3020 int i, d;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3021 for( i = 0; i < 4; i++ ) {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3022 if( tc0[i] < 0 ) {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3023 pix += 4*ystride;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3024 continue;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3025 }
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3026 for( d = 0; d < 4; d++ ) {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3027 const int p0 = pix[-1*xstride];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3028 const int p1 = pix[-2*xstride];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3029 const int p2 = pix[-3*xstride];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3030 const int q0 = pix[0];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3031 const int q1 = pix[1*xstride];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3032 const int q2 = pix[2*xstride];
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3033
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3034 if( FFABS( p0 - q0 ) < alpha &&
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3035 FFABS( p1 - p0 ) < beta &&
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3036 FFABS( q1 - q0 ) < beta ) {
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3037
2633
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3038 int tc = tc0[i];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3039 int i_delta;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3040
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3041 if( FFABS( p2 - p0 ) < beta ) {
10940
563cb9b1a9b7 skip outer pixels if possible in h264_loop_filter_luma_c().
michael
parents: 10878
diff changeset
3042 if(tc0[i])
4594
a96d905dcbaa Add av_ prefix to clip functions
reimar
parents: 4296
diff changeset
3043 pix[-2*xstride] = p1 + av_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
3044 tc++;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3045 }
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3046 if( FFABS( q2 - q0 ) < beta ) {
10940
563cb9b1a9b7 skip outer pixels if possible in h264_loop_filter_luma_c().
michael
parents: 10878
diff changeset
3047 if(tc0[i])
4594
a96d905dcbaa Add av_ prefix to clip functions
reimar
parents: 4296
diff changeset
3048 pix[ xstride] = q1 + av_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
3049 tc++;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3050 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3051
4594
a96d905dcbaa Add av_ prefix to clip functions
reimar
parents: 4296
diff changeset
3052 i_delta = av_clip( (((q0 - p0 ) << 2) + (p1 - q1) + 4) >> 3, -tc, tc );
a96d905dcbaa Add av_ prefix to clip functions
reimar
parents: 4296
diff changeset
3053 pix[-xstride] = av_clip_uint8( p0 + i_delta ); /* p0' */
a96d905dcbaa Add av_ prefix to clip functions
reimar
parents: 4296
diff changeset
3054 pix[0] = av_clip_uint8( q0 - i_delta ); /* q0' */
2633
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3055 }
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3056 pix += ystride;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3057 }
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3058 }
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3059 }
2707
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
3060 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
3061 {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3062 h264_loop_filter_luma_c(pix, stride, 1, alpha, beta, tc0);
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3063 }
2707
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
3064 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
3065 {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3066 h264_loop_filter_luma_c(pix, 1, stride, alpha, beta, tc0);
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3067 }
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3068
10941
28edcc8c54c0 Mark the h264 c loop filter as av_always_inline av_flatten to make sure its
michael
parents: 10940
diff changeset
3069 static av_always_inline av_flatten void h264_loop_filter_luma_intra_c(uint8_t *pix, int xstride, int ystride, int alpha, int beta)
8395
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3070 {
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3071 int d;
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3072 for( d = 0; d < 16; d++ ) {
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3073 const int p2 = pix[-3*xstride];
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3074 const int p1 = pix[-2*xstride];
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3075 const int p0 = pix[-1*xstride];
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3076
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3077 const int q0 = pix[ 0*xstride];
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3078 const int q1 = pix[ 1*xstride];
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3079 const int q2 = pix[ 2*xstride];
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3080
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3081 if( FFABS( p0 - q0 ) < alpha &&
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3082 FFABS( p1 - p0 ) < beta &&
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3083 FFABS( q1 - q0 ) < beta ) {
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3084
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3085 if(FFABS( p0 - q0 ) < (( alpha >> 2 ) + 2 )){
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3086 if( FFABS( p2 - p0 ) < beta)
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3087 {
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3088 const int p3 = pix[-4*xstride];
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3089 /* p0', p1', p2' */
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3090 pix[-1*xstride] = ( p2 + 2*p1 + 2*p0 + 2*q0 + q1 + 4 ) >> 3;
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3091 pix[-2*xstride] = ( p2 + p1 + p0 + q0 + 2 ) >> 2;
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3092 pix[-3*xstride] = ( 2*p3 + 3*p2 + p1 + p0 + q0 + 4 ) >> 3;
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3093 } else {
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3094 /* p0' */
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3095 pix[-1*xstride] = ( 2*p1 + p0 + q1 + 2 ) >> 2;
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3096 }
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3097 if( FFABS( q2 - q0 ) < beta)
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3098 {
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3099 const int q3 = pix[3*xstride];
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3100 /* q0', q1', q2' */
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3101 pix[0*xstride] = ( p1 + 2*p0 + 2*q0 + 2*q1 + q2 + 4 ) >> 3;
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3102 pix[1*xstride] = ( p0 + q0 + q1 + q2 + 2 ) >> 2;
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3103 pix[2*xstride] = ( 2*q3 + 3*q2 + q1 + q0 + p0 + 4 ) >> 3;
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3104 } else {
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3105 /* q0' */
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3106 pix[0*xstride] = ( 2*q1 + q0 + p1 + 2 ) >> 2;
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3107 }
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3108 }else{
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3109 /* p0', q0' */
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3110 pix[-1*xstride] = ( 2*p1 + p0 + q1 + 2 ) >> 2;
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3111 pix[ 0*xstride] = ( 2*q1 + q0 + p1 + 2 ) >> 2;
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3112 }
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3113 }
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3114 pix += ystride;
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3115 }
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3116 }
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3117 static void h264_v_loop_filter_luma_intra_c(uint8_t *pix, int stride, int alpha, int beta)
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3118 {
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3119 h264_loop_filter_luma_intra_c(pix, stride, 1, alpha, beta);
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3120 }
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3121 static void h264_h_loop_filter_luma_intra_c(uint8_t *pix, int stride, int alpha, int beta)
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3122 {
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3123 h264_loop_filter_luma_intra_c(pix, 1, stride, alpha, beta);
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3124 }
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
3125
10941
28edcc8c54c0 Mark the h264 c loop filter as av_always_inline av_flatten to make sure its
michael
parents: 10940
diff changeset
3126 static av_always_inline av_flatten 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
3127 {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3128 int i, d;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3129 for( i = 0; i < 4; i++ ) {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3130 const int tc = tc0[i];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3131 if( tc <= 0 ) {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3132 pix += 2*ystride;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3133 continue;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3134 }
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3135 for( d = 0; d < 2; d++ ) {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3136 const int p0 = pix[-1*xstride];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3137 const int p1 = pix[-2*xstride];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3138 const int q0 = pix[0];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3139 const int q1 = pix[1*xstride];
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3140
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3141 if( FFABS( p0 - q0 ) < alpha &&
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3142 FFABS( p1 - p0 ) < beta &&
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3143 FFABS( q1 - q0 ) < beta ) {
2633
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3144
4594
a96d905dcbaa Add av_ prefix to clip functions
reimar
parents: 4296
diff changeset
3145 int delta = av_clip( (((q0 - p0 ) << 2) + (p1 - q1) + 4) >> 3, -tc, tc );
a96d905dcbaa Add av_ prefix to clip functions
reimar
parents: 4296
diff changeset
3146
a96d905dcbaa Add av_ prefix to clip functions
reimar
parents: 4296
diff changeset
3147 pix[-xstride] = av_clip_uint8( p0 + delta ); /* p0' */
a96d905dcbaa Add av_ prefix to clip functions
reimar
parents: 4296
diff changeset
3148 pix[0] = av_clip_uint8( q0 - delta ); /* q0' */
2633
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3149 }
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3150 pix += ystride;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3151 }
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3152 }
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3153 }
2707
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
3154 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
3155 {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3156 h264_loop_filter_chroma_c(pix, stride, 1, alpha, beta, tc0);
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3157 }
2707
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
3158 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
3159 {
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3160 h264_loop_filter_chroma_c(pix, 1, stride, alpha, beta, tc0);
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3161 }
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
3162
10941
28edcc8c54c0 Mark the h264 c loop filter as av_always_inline av_flatten to make sure its
michael
parents: 10940
diff changeset
3163 static av_always_inline av_flatten void h264_loop_filter_chroma_intra_c(uint8_t *pix, int xstride, int ystride, int alpha, int beta)
2707
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
3164 {
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
3165 int d;
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
3166 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
3167 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
3168 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
3169 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
3170 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
3171
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3172 if( FFABS( p0 - q0 ) < alpha &&
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3173 FFABS( p1 - p0 ) < beta &&
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3174 FFABS( q1 - q0 ) < beta ) {
2707
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
3175
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
3176 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
3177 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
3178 }
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
3179 pix += ystride;
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
3180 }
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
3181 }
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
3182 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
3183 {
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
3184 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
3185 }
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
3186 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
3187 {
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
3188 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
3189 }
360024d31dab H.264 deblocking optimizations (mmx for chroma_bS4 case, convert existing cases to 8-bit math)
lorenm
parents: 2696
diff changeset
3190
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3191 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
3192 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
3193 int s, i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3194
986e461dc072 Initial revision
glantau
parents:
diff changeset
3195 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3196 for(i=0;i<h;i++) {
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
3197 s += abs(pix1[0] - pix2[0]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
3198 s += abs(pix1[1] - pix2[1]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
3199 s += abs(pix1[2] - pix2[2]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
3200 s += abs(pix1[3] - pix2[3]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
3201 s += abs(pix1[4] - pix2[4]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
3202 s += abs(pix1[5] - pix2[5]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
3203 s += abs(pix1[6] - pix2[6]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
3204 s += abs(pix1[7] - pix2[7]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
3205 s += abs(pix1[8] - pix2[8]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
3206 s += abs(pix1[9] - pix2[9]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
3207 s += abs(pix1[10] - pix2[10]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
3208 s += abs(pix1[11] - pix2[11]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
3209 s += abs(pix1[12] - pix2[12]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
3210 s += abs(pix1[13] - pix2[13]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
3211 s += abs(pix1[14] - pix2[14]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
3212 s += abs(pix1[15] - pix2[15]);
986e461dc072 Initial revision
glantau
parents:
diff changeset
3213 pix1 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3214 pix2 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3215 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
3216 return s;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3217 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
3218
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3219 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
3220 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
3221 int s, i;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3222
986e461dc072 Initial revision
glantau
parents:
diff changeset
3223 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3224 for(i=0;i<h;i++) {
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
3225 s += abs(pix1[0] - avg2(pix2[0], pix2[1]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3226 s += abs(pix1[1] - avg2(pix2[1], pix2[2]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3227 s += abs(pix1[2] - avg2(pix2[2], pix2[3]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3228 s += abs(pix1[3] - avg2(pix2[3], pix2[4]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3229 s += abs(pix1[4] - avg2(pix2[4], pix2[5]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3230 s += abs(pix1[5] - avg2(pix2[5], pix2[6]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3231 s += abs(pix1[6] - avg2(pix2[6], pix2[7]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3232 s += abs(pix1[7] - avg2(pix2[7], pix2[8]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3233 s += abs(pix1[8] - avg2(pix2[8], pix2[9]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3234 s += abs(pix1[9] - avg2(pix2[9], pix2[10]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3235 s += abs(pix1[10] - avg2(pix2[10], pix2[11]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3236 s += abs(pix1[11] - avg2(pix2[11], pix2[12]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3237 s += abs(pix1[12] - avg2(pix2[12], pix2[13]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3238 s += abs(pix1[13] - avg2(pix2[13], pix2[14]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3239 s += abs(pix1[14] - avg2(pix2[14], pix2[15]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3240 s += abs(pix1[15] - avg2(pix2[15], pix2[16]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3241 pix1 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3242 pix2 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3243 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
3244 return s;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3245 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
3246
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3247 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
3248 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
3249 int s, i;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
3250 uint8_t *pix3 = pix2 + line_size;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
3251
986e461dc072 Initial revision
glantau
parents:
diff changeset
3252 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3253 for(i=0;i<h;i++) {
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
3254 s += abs(pix1[0] - avg2(pix2[0], pix3[0]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3255 s += abs(pix1[1] - avg2(pix2[1], pix3[1]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3256 s += abs(pix1[2] - avg2(pix2[2], pix3[2]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3257 s += abs(pix1[3] - avg2(pix2[3], pix3[3]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3258 s += abs(pix1[4] - avg2(pix2[4], pix3[4]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3259 s += abs(pix1[5] - avg2(pix2[5], pix3[5]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3260 s += abs(pix1[6] - avg2(pix2[6], pix3[6]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3261 s += abs(pix1[7] - avg2(pix2[7], pix3[7]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3262 s += abs(pix1[8] - avg2(pix2[8], pix3[8]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3263 s += abs(pix1[9] - avg2(pix2[9], pix3[9]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3264 s += abs(pix1[10] - avg2(pix2[10], pix3[10]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3265 s += abs(pix1[11] - avg2(pix2[11], pix3[11]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3266 s += abs(pix1[12] - avg2(pix2[12], pix3[12]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3267 s += abs(pix1[13] - avg2(pix2[13], pix3[13]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3268 s += abs(pix1[14] - avg2(pix2[14], pix3[14]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3269 s += abs(pix1[15] - avg2(pix2[15], pix3[15]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3270 pix1 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3271 pix2 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3272 pix3 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3273 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
3274 return s;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3275 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
3276
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3277 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
3278 {
986e461dc072 Initial revision
glantau
parents:
diff changeset
3279 int s, i;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
3280 uint8_t *pix3 = pix2 + line_size;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
3281
986e461dc072 Initial revision
glantau
parents:
diff changeset
3282 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3283 for(i=0;i<h;i++) {
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
3284 s += abs(pix1[0] - avg4(pix2[0], pix2[1], pix3[0], pix3[1]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3285 s += abs(pix1[1] - avg4(pix2[1], pix2[2], pix3[1], pix3[2]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3286 s += abs(pix1[2] - avg4(pix2[2], pix2[3], pix3[2], pix3[3]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3287 s += abs(pix1[3] - avg4(pix2[3], pix2[4], pix3[3], pix3[4]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3288 s += abs(pix1[4] - avg4(pix2[4], pix2[5], pix3[4], pix3[5]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3289 s += abs(pix1[5] - avg4(pix2[5], pix2[6], pix3[5], pix3[6]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3290 s += abs(pix1[6] - avg4(pix2[6], pix2[7], pix3[6], pix3[7]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3291 s += abs(pix1[7] - avg4(pix2[7], pix2[8], pix3[7], pix3[8]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3292 s += abs(pix1[8] - avg4(pix2[8], pix2[9], pix3[8], pix3[9]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3293 s += abs(pix1[9] - avg4(pix2[9], pix2[10], pix3[9], pix3[10]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3294 s += abs(pix1[10] - avg4(pix2[10], pix2[11], pix3[10], pix3[11]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3295 s += abs(pix1[11] - avg4(pix2[11], pix2[12], pix3[11], pix3[12]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3296 s += abs(pix1[12] - avg4(pix2[12], pix2[13], pix3[12], pix3[13]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3297 s += abs(pix1[13] - avg4(pix2[13], pix2[14], pix3[13], pix3[14]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3298 s += abs(pix1[14] - avg4(pix2[14], pix2[15], pix3[14], pix3[15]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3299 s += abs(pix1[15] - avg4(pix2[15], pix2[16], pix3[15], pix3[16]));
986e461dc072 Initial revision
glantau
parents:
diff changeset
3300 pix1 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3301 pix2 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3302 pix3 += line_size;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3303 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
3304 return s;
986e461dc072 Initial revision
glantau
parents:
diff changeset
3305 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
3306
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3307 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
3308 {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3309 int s, i;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3310
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3311 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3312 for(i=0;i<h;i++) {
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3313 s += abs(pix1[0] - pix2[0]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3314 s += abs(pix1[1] - pix2[1]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3315 s += abs(pix1[2] - pix2[2]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3316 s += abs(pix1[3] - pix2[3]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3317 s += abs(pix1[4] - pix2[4]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3318 s += abs(pix1[5] - pix2[5]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3319 s += abs(pix1[6] - pix2[6]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3320 s += abs(pix1[7] - pix2[7]);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3321 pix1 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3322 pix2 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3323 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3324 return s;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3325 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3326
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3327 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
3328 {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3329 int s, i;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3330
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3331 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3332 for(i=0;i<h;i++) {
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3333 s += abs(pix1[0] - avg2(pix2[0], pix2[1]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3334 s += abs(pix1[1] - avg2(pix2[1], pix2[2]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3335 s += abs(pix1[2] - avg2(pix2[2], pix2[3]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3336 s += abs(pix1[3] - avg2(pix2[3], pix2[4]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3337 s += abs(pix1[4] - avg2(pix2[4], pix2[5]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3338 s += abs(pix1[5] - avg2(pix2[5], pix2[6]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3339 s += abs(pix1[6] - avg2(pix2[6], pix2[7]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3340 s += abs(pix1[7] - avg2(pix2[7], pix2[8]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3341 pix1 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3342 pix2 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3343 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3344 return s;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3345 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3346
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3347 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
3348 {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3349 int s, i;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
3350 uint8_t *pix3 = pix2 + line_size;
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3351
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3352 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3353 for(i=0;i<h;i++) {
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3354 s += abs(pix1[0] - avg2(pix2[0], pix3[0]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3355 s += abs(pix1[1] - avg2(pix2[1], pix3[1]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3356 s += abs(pix1[2] - avg2(pix2[2], pix3[2]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3357 s += abs(pix1[3] - avg2(pix2[3], pix3[3]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3358 s += abs(pix1[4] - avg2(pix2[4], pix3[4]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3359 s += abs(pix1[5] - avg2(pix2[5], pix3[5]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3360 s += abs(pix1[6] - avg2(pix2[6], pix3[6]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3361 s += abs(pix1[7] - avg2(pix2[7], pix3[7]));
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3362 pix1 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3363 pix2 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3364 pix3 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3365 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3366 return s;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3367 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3368
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3369 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
3370 {
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3371 int s, i;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
3372 uint8_t *pix3 = pix2 + line_size;
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3373
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3374 s = 0;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3375 for(i=0;i<h;i++) {
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3376 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
3377 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
3378 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
3379 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
3380 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
3381 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
3382 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
3383 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
3384 pix1 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3385 pix2 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3386 pix3 += line_size;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3387 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3388 return s;
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3389 }
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 256
diff changeset
3390
2834
fd5d7c732c6b kill a bunch of compiler warnings
mru
parents: 2763
diff changeset
3391 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
3392 MpegEncContext *c = v;
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3393 int score1=0;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3394 int score2=0;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3395 int x,y;
2066
4bfb146e701b nsse weight
michael
parents: 2065
diff changeset
3396
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3397 for(y=0; y<h; y++){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3398 for(x=0; x<16; x++){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3399 score1+= (s1[x ] - s2[x ])*(s1[x ] - s2[x ]);
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3400 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3401 if(y+1<h){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3402 for(x=0; x<15; x++){
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3403 score2+= FFABS( s1[x ] - s1[x +stride]
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3404 - s1[x+1] + s1[x+1+stride])
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3405 -FFABS( s2[x ] - s2[x +stride]
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3406 - s2[x+1] + s2[x+1+stride]);
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3407 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3408 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3409 s1+= stride;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3410 s2+= stride;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3411 }
2066
4bfb146e701b nsse weight
michael
parents: 2065
diff changeset
3412
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3413 if(c) return score1 + FFABS(score2)*c->avctx->nsse_weight;
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3414 else return score1 + FFABS(score2)*8;
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3415 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3416
2834
fd5d7c732c6b kill a bunch of compiler warnings
mru
parents: 2763
diff changeset
3417 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
3418 MpegEncContext *c = v;
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3419 int score1=0;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3420 int score2=0;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3421 int x,y;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3422
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3423 for(y=0; y<h; y++){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3424 for(x=0; x<8; x++){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3425 score1+= (s1[x ] - s2[x ])*(s1[x ] - s2[x ]);
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3426 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3427 if(y+1<h){
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3428 for(x=0; x<7; x++){
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3429 score2+= FFABS( s1[x ] - s1[x +stride]
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3430 - s1[x+1] + s1[x+1+stride])
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3431 -FFABS( s2[x ] - s2[x +stride]
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3432 - s2[x+1] + s2[x+1+stride]);
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3433 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3434 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3435 s1+= stride;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3436 s2+= stride;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3437 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3438
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3439 if(c) return score1 + FFABS(score2)*c->avctx->nsse_weight;
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3440 else return score1 + FFABS(score2)*8;
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3441 }
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3442
1784
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3443 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
3444 int i;
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3445 unsigned int sum=0;
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3446
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3447 for(i=0; i<8*8; i++){
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3448 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
3449 int w= weight[i];
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3450 b>>= RECON_SHIFT;
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3451 assert(-512<b && b<512);
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3452
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3453 sum += (w*b)*(w*b)>>4;
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3454 }
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3455 return sum>>2;
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3456 }
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3457
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3458 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
3459 int i;
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3460
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3461 for(i=0; i<8*8; i++){
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3462 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
3463 }
1784
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3464 }
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
3465
1100
207f27932628 ff_block_permute() dox
michaelni
parents: 1092
diff changeset
3466 /**
207f27932628 ff_block_permute() dox
michaelni
parents: 1092
diff changeset
3467 * permutes an 8x8 block.
1101
a3a8017dee70 more idiot proof dox
michaelni
parents: 1100
diff changeset
3468 * @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
3469 * @param permutation the permutation vector
207f27932628 ff_block_permute() dox
michaelni
parents: 1092
diff changeset
3470 * @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
3471 * @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
3472 * (inverse) permutated to scantable order!
1100
207f27932628 ff_block_permute() dox
michaelni
parents: 1092
diff changeset
3473 */
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
3474 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
3475 {
764
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3476 int i;
945
463f7260b155 trellis quantization
michaelni
parents: 936
diff changeset
3477 DCTELEM temp[64];
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3478
764
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3479 if(last<=0) return;
5129
0244bba24b43 misc typo fixes
diego
parents: 5065
diff changeset
3480 //if(permutation[1]==1) return; //FIXME it is ok but not clean and might fail for some permutations
174
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
3481
764
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3482 for(i=0; i<=last; i++){
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3483 const int j= scantable[i];
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3484 temp[j]= block[j];
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3485 block[j]=0;
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3486 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3487
764
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3488 for(i=0; i<=last; i++){
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3489 const int j= scantable[i];
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3490 const int perm_j= permutation[j];
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3491 block[perm_j]= temp[j];
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
3492 }
174
ac5075a55488 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents: 88
diff changeset
3493 }
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
3494
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3495 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
3496 return 0;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3497 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3498
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3499 void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3500 int i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3501
8976
e7d87561b42b Making the arrays accomodate an extra intra 8x8 cmp function
romansh
parents: 8785
diff changeset
3502 memset(cmp, 0, sizeof(void*)*6);
e7d87561b42b Making the arrays accomodate an extra intra 8x8 cmp function
romansh
parents: 8785
diff changeset
3503
e7d87561b42b Making the arrays accomodate an extra intra 8x8 cmp function
romansh
parents: 8785
diff changeset
3504 for(i=0; i<6; i++){
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3505 switch(type&0xFF){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3506 case FF_CMP_SAD:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3507 cmp[i]= c->sad[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3508 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3509 case FF_CMP_SATD:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3510 cmp[i]= c->hadamard8_diff[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3511 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3512 case FF_CMP_SSE:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3513 cmp[i]= c->sse[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3514 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3515 case FF_CMP_DCT:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3516 cmp[i]= c->dct_sad[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3517 break;
3010
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3518 case FF_CMP_DCT264:
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3519 cmp[i]= c->dct264_sad[i];
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3520 break;
2382
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3521 case FF_CMP_DCTMAX:
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3522 cmp[i]= c->dct_max[i];
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3523 break;
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3524 case FF_CMP_PSNR:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3525 cmp[i]= c->quant_psnr[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3526 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3527 case FF_CMP_BIT:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3528 cmp[i]= c->bit[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3529 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3530 case FF_CMP_RD:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3531 cmp[i]= c->rd[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3532 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3533 case FF_CMP_VSAD:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3534 cmp[i]= c->vsad[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3535 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3536 case FF_CMP_VSSE:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3537 cmp[i]= c->vsse[i];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3538 break;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3539 case FF_CMP_ZERO:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3540 cmp[i]= zero_cmp;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3541 break;
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3542 case FF_CMP_NSSE:
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3543 cmp[i]= c->nsse[i];
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
3544 break;
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8496
diff changeset
3545 #if CONFIG_SNOW_ENCODER
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
3546 case FF_CMP_W53:
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
3547 cmp[i]= c->w53[i];
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
3548 break;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
3549 case FF_CMP_W97:
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
3550 cmp[i]= c->w97[i];
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
3551 break;
3373
b8996cc5ccae Disable w53 and w97 cmp methods when snow encoder is disabled
gpoirier
parents: 3323
diff changeset
3552 #endif
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3553 default:
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3554 av_log(NULL, AV_LOG_ERROR,"internal error in cmp function selection\n");
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3555 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3556 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3557 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3558
8288
800444234375 clear_block mmx
lorenm
parents: 8232
diff changeset
3559 static void clear_block_c(DCTELEM *block)
800444234375 clear_block mmx
lorenm
parents: 8232
diff changeset
3560 {
800444234375 clear_block mmx
lorenm
parents: 8232
diff changeset
3561 memset(block, 0, sizeof(DCTELEM)*64);
800444234375 clear_block mmx
lorenm
parents: 8232
diff changeset
3562 }
800444234375 clear_block mmx
lorenm
parents: 8232
diff changeset
3563
1101
a3a8017dee70 more idiot proof dox
michaelni
parents: 1100
diff changeset
3564 /**
a3a8017dee70 more idiot proof dox
michaelni
parents: 1100
diff changeset
3565 * memset(blocks, 0, sizeof(DCTELEM)*6*64)
a3a8017dee70 more idiot proof dox
michaelni
parents: 1100
diff changeset
3566 */
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
3567 static void clear_blocks_c(DCTELEM *blocks)
296
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
3568 {
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
3569 memset(blocks, 0, sizeof(DCTELEM)*6*64);
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
3570 }
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
3571
866
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3572 static void add_bytes_c(uint8_t *dst, uint8_t *src, int w){
6385
40fbc878ce3f pseudo-simd add_bytes and diff_bytes
lorenm
parents: 6384
diff changeset
3573 long i;
40fbc878ce3f pseudo-simd add_bytes and diff_bytes
lorenm
parents: 6384
diff changeset
3574 for(i=0; i<=w-sizeof(long); i+=sizeof(long)){
40fbc878ce3f pseudo-simd add_bytes and diff_bytes
lorenm
parents: 6384
diff changeset
3575 long a = *(long*)(src+i);
40fbc878ce3f pseudo-simd add_bytes and diff_bytes
lorenm
parents: 6384
diff changeset
3576 long b = *(long*)(dst+i);
40fbc878ce3f pseudo-simd add_bytes and diff_bytes
lorenm
parents: 6384
diff changeset
3577 *(long*)(dst+i) = ((a&pb_7f) + (b&pb_7f)) ^ ((a^b)&pb_80);
866
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3578 }
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3579 for(; i<w; i++)
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3580 dst[i+0] += src[i+0];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3581 }
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3582
6384
0a403ade8c81 simd and unroll png_filter_row
lorenm
parents: 6241
diff changeset
3583 static void add_bytes_l2_c(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w){
6385
40fbc878ce3f pseudo-simd add_bytes and diff_bytes
lorenm
parents: 6384
diff changeset
3584 long i;
6384
0a403ade8c81 simd and unroll png_filter_row
lorenm
parents: 6241
diff changeset
3585 for(i=0; i<=w-sizeof(long); i+=sizeof(long)){
0a403ade8c81 simd and unroll png_filter_row
lorenm
parents: 6241
diff changeset
3586 long a = *(long*)(src1+i);
0a403ade8c81 simd and unroll png_filter_row
lorenm
parents: 6241
diff changeset
3587 long b = *(long*)(src2+i);
6385
40fbc878ce3f pseudo-simd add_bytes and diff_bytes
lorenm
parents: 6384
diff changeset
3588 *(long*)(dst+i) = ((a&pb_7f) + (b&pb_7f)) ^ ((a^b)&pb_80);
6384
0a403ade8c81 simd and unroll png_filter_row
lorenm
parents: 6241
diff changeset
3589 }
0a403ade8c81 simd and unroll png_filter_row
lorenm
parents: 6241
diff changeset
3590 for(; i<w; i++)
0a403ade8c81 simd and unroll png_filter_row
lorenm
parents: 6241
diff changeset
3591 dst[i] = src1[i]+src2[i];
0a403ade8c81 simd and unroll png_filter_row
lorenm
parents: 6241
diff changeset
3592 }
0a403ade8c81 simd and unroll png_filter_row
lorenm
parents: 6241
diff changeset
3593
866
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3594 static void diff_bytes_c(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w){
6385
40fbc878ce3f pseudo-simd add_bytes and diff_bytes
lorenm
parents: 6384
diff changeset
3595 long i;
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8496
diff changeset
3596 #if !HAVE_FAST_UNALIGNED
6385
40fbc878ce3f pseudo-simd add_bytes and diff_bytes
lorenm
parents: 6384
diff changeset
3597 if((long)src2 & (sizeof(long)-1)){
6386
lorenm
parents: 6385
diff changeset
3598 for(i=0; i+7<w; i+=8){
lorenm
parents: 6385
diff changeset
3599 dst[i+0] = src1[i+0]-src2[i+0];
lorenm
parents: 6385
diff changeset
3600 dst[i+1] = src1[i+1]-src2[i+1];
lorenm
parents: 6385
diff changeset
3601 dst[i+2] = src1[i+2]-src2[i+2];
lorenm
parents: 6385
diff changeset
3602 dst[i+3] = src1[i+3]-src2[i+3];
lorenm
parents: 6385
diff changeset
3603 dst[i+4] = src1[i+4]-src2[i+4];
lorenm
parents: 6385
diff changeset
3604 dst[i+5] = src1[i+5]-src2[i+5];
lorenm
parents: 6385
diff changeset
3605 dst[i+6] = src1[i+6]-src2[i+6];
lorenm
parents: 6385
diff changeset
3606 dst[i+7] = src1[i+7]-src2[i+7];
lorenm
parents: 6385
diff changeset
3607 }
6385
40fbc878ce3f pseudo-simd add_bytes and diff_bytes
lorenm
parents: 6384
diff changeset
3608 }else
40fbc878ce3f pseudo-simd add_bytes and diff_bytes
lorenm
parents: 6384
diff changeset
3609 #endif
40fbc878ce3f pseudo-simd add_bytes and diff_bytes
lorenm
parents: 6384
diff changeset
3610 for(i=0; i<=w-sizeof(long); i+=sizeof(long)){
40fbc878ce3f pseudo-simd add_bytes and diff_bytes
lorenm
parents: 6384
diff changeset
3611 long a = *(long*)(src1+i);
40fbc878ce3f pseudo-simd add_bytes and diff_bytes
lorenm
parents: 6384
diff changeset
3612 long b = *(long*)(src2+i);
40fbc878ce3f pseudo-simd add_bytes and diff_bytes
lorenm
parents: 6384
diff changeset
3613 *(long*)(dst+i) = ((a|pb_80) - (b&pb_7f)) ^ ((a^b^pb_80)&pb_80);
40fbc878ce3f pseudo-simd add_bytes and diff_bytes
lorenm
parents: 6384
diff changeset
3614 }
866
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3615 for(; i<w; i++)
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3616 dst[i+0] = src1[i+0]-src2[i+0];
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3617 }
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
3618
10431
546b7ebeaf07 huffyuv: add some const qualifiers
lorenm
parents: 10429
diff changeset
3619 static void add_hfyu_median_prediction_c(uint8_t *dst, const uint8_t *src1, const uint8_t *diff, int w, int *left, int *left_top){
8760
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8719
diff changeset
3620 int i;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8719
diff changeset
3621 uint8_t l, lt;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8719
diff changeset
3622
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8719
diff changeset
3623 l= *left;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8719
diff changeset
3624 lt= *left_top;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8719
diff changeset
3625
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8719
diff changeset
3626 for(i=0; i<w; i++){
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8719
diff changeset
3627 l= mid_pred(l, src1[i], (l + src1[i] - lt)&0xFF) + diff[i];
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8719
diff changeset
3628 lt= src1[i];
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8719
diff changeset
3629 dst[i]= l;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8719
diff changeset
3630 }
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8719
diff changeset
3631
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8719
diff changeset
3632 *left= l;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8719
diff changeset
3633 *left_top= lt;
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8719
diff changeset
3634 }
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8719
diff changeset
3635
10431
546b7ebeaf07 huffyuv: add some const qualifiers
lorenm
parents: 10429
diff changeset
3636 static void sub_hfyu_median_prediction_c(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int w, int *left, int *left_top){
1527
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3637 int i;
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3638 uint8_t l, lt;
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3639
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3640 l= *left;
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3641 lt= *left_top;
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3642
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3643 for(i=0; i<w; i++){
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3644 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
3645 lt= src1[i];
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3646 l= src2[i];
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3647 dst[i]= l - pred;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3648 }
1527
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3649
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3650 *left= l;
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3651 *left_top= lt;
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3652 }
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
3653
10420
442ab0c41eae Huffyuv: Add missing const to src pointers in dsputil functions.
astrange
parents: 10370
diff changeset
3654 static int add_hfyu_left_prediction_c(uint8_t *dst, const uint8_t *src, int w, int acc){
10370
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3655 int i;
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3656
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3657 for(i=0; i<w-1; i++){
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3658 acc+= src[i];
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3659 dst[i]= acc;
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3660 i++;
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3661 acc+= src[i];
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3662 dst[i]= acc;
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3663 }
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3664
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3665 for(; i<w; i++){
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3666 acc+= src[i];
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3667 dst[i]= acc;
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3668 }
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3669
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3670 return acc;
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3671 }
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3672
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3673 #if HAVE_BIGENDIAN
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3674 #define B 3
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3675 #define G 2
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3676 #define R 1
10878
a8620b001ed3 Implement alpha channel decoding for BGR HuffYUV.
astrange
parents: 10867
diff changeset
3677 #define A 0
10370
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3678 #else
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3679 #define B 0
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3680 #define G 1
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3681 #define R 2
10878
a8620b001ed3 Implement alpha channel decoding for BGR HuffYUV.
astrange
parents: 10867
diff changeset
3682 #define A 3
10370
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3683 #endif
10878
a8620b001ed3 Implement alpha channel decoding for BGR HuffYUV.
astrange
parents: 10867
diff changeset
3684 static void add_hfyu_left_prediction_bgr32_c(uint8_t *dst, const uint8_t *src, int w, int *red, int *green, int *blue, int *alpha){
10370
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3685 int i;
10878
a8620b001ed3 Implement alpha channel decoding for BGR HuffYUV.
astrange
parents: 10867
diff changeset
3686 int r,g,b,a;
10370
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3687 r= *red;
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3688 g= *green;
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3689 b= *blue;
10878
a8620b001ed3 Implement alpha channel decoding for BGR HuffYUV.
astrange
parents: 10867
diff changeset
3690 a= *alpha;
10370
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3691
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3692 for(i=0; i<w; i++){
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3693 b+= src[4*i+B];
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3694 g+= src[4*i+G];
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3695 r+= src[4*i+R];
10878
a8620b001ed3 Implement alpha channel decoding for BGR HuffYUV.
astrange
parents: 10867
diff changeset
3696 a+= src[4*i+A];
10370
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3697
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3698 dst[4*i+B]= b;
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3699 dst[4*i+G]= g;
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3700 dst[4*i+R]= r;
10878
a8620b001ed3 Implement alpha channel decoding for BGR HuffYUV.
astrange
parents: 10867
diff changeset
3701 dst[4*i+A]= a;
10370
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3702 }
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3703
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3704 *red= r;
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3705 *green= g;
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3706 *blue= b;
10878
a8620b001ed3 Implement alpha channel decoding for BGR HuffYUV.
astrange
parents: 10867
diff changeset
3707 *alpha= a;
10370
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3708 }
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3709 #undef B
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3710 #undef G
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3711 #undef R
10878
a8620b001ed3 Implement alpha channel decoding for BGR HuffYUV.
astrange
parents: 10867
diff changeset
3712 #undef A
10370
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
3713
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3714 #define BUTTERFLY2(o1,o2,i1,i2) \
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3715 o1= (i1)+(i2);\
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3716 o2= (i1)-(i2);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3717
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3718 #define BUTTERFLY1(x,y) \
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3719 {\
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3720 int a,b;\
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3721 a= x;\
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3722 b= y;\
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3723 x= a+b;\
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3724 y= a-b;\
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3725 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3726
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3727 #define BUTTERFLYA(x,y) (FFABS((x)+(y)) + FFABS((x)-(y)))
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3728
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3729 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
3730 int i;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3731 int temp[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3732 int sum=0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3733
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3734 assert(h==8);
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3735
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3736 for(i=0; i<8; i++){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3737 //FIXME try pointer walks
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3738 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
3739 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
3740 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
3741 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
3742
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3743 BUTTERFLY1(temp[8*i+0], temp[8*i+2]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3744 BUTTERFLY1(temp[8*i+1], temp[8*i+3]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3745 BUTTERFLY1(temp[8*i+4], temp[8*i+6]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3746 BUTTERFLY1(temp[8*i+5], temp[8*i+7]);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3747
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3748 BUTTERFLY1(temp[8*i+0], temp[8*i+4]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3749 BUTTERFLY1(temp[8*i+1], temp[8*i+5]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3750 BUTTERFLY1(temp[8*i+2], temp[8*i+6]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3751 BUTTERFLY1(temp[8*i+3], temp[8*i+7]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3752 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3753
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3754 for(i=0; i<8; i++){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3755 BUTTERFLY1(temp[8*0+i], temp[8*1+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3756 BUTTERFLY1(temp[8*2+i], temp[8*3+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3757 BUTTERFLY1(temp[8*4+i], temp[8*5+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3758 BUTTERFLY1(temp[8*6+i], temp[8*7+i]);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3759
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3760 BUTTERFLY1(temp[8*0+i], temp[8*2+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3761 BUTTERFLY1(temp[8*1+i], temp[8*3+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3762 BUTTERFLY1(temp[8*4+i], temp[8*6+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3763 BUTTERFLY1(temp[8*5+i], temp[8*7+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3764
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3765 sum +=
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3766 BUTTERFLYA(temp[8*0+i], temp[8*4+i])
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3767 +BUTTERFLYA(temp[8*1+i], temp[8*5+i])
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3768 +BUTTERFLYA(temp[8*2+i], temp[8*6+i])
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3769 +BUTTERFLYA(temp[8*3+i], temp[8*7+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3770 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3771 #if 0
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3772 static int maxi=0;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3773 if(sum>maxi){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3774 maxi=sum;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3775 printf("MAX:%d\n", maxi);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3776 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3777 #endif
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3778 return sum;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3779 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3780
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3781 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
3782 int i;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3783 int temp[64];
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3784 int sum=0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3785
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3786 assert(h==8);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3787
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3788 for(i=0; i<8; i++){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3789 //FIXME try pointer walks
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
3790 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
3791 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
3792 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
3793 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
3794
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3795 BUTTERFLY1(temp[8*i+0], temp[8*i+2]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3796 BUTTERFLY1(temp[8*i+1], temp[8*i+3]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3797 BUTTERFLY1(temp[8*i+4], temp[8*i+6]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3798 BUTTERFLY1(temp[8*i+5], temp[8*i+7]);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3799
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3800 BUTTERFLY1(temp[8*i+0], temp[8*i+4]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3801 BUTTERFLY1(temp[8*i+1], temp[8*i+5]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3802 BUTTERFLY1(temp[8*i+2], temp[8*i+6]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3803 BUTTERFLY1(temp[8*i+3], temp[8*i+7]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3804 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3805
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3806 for(i=0; i<8; i++){
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3807 BUTTERFLY1(temp[8*0+i], temp[8*1+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3808 BUTTERFLY1(temp[8*2+i], temp[8*3+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3809 BUTTERFLY1(temp[8*4+i], temp[8*5+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3810 BUTTERFLY1(temp[8*6+i], temp[8*7+i]);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3811
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3812 BUTTERFLY1(temp[8*0+i], temp[8*2+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3813 BUTTERFLY1(temp[8*1+i], temp[8*3+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3814 BUTTERFLY1(temp[8*4+i], temp[8*6+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3815 BUTTERFLY1(temp[8*5+i], temp[8*7+i]);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3816
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3817 sum +=
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3818 BUTTERFLYA(temp[8*0+i], temp[8*4+i])
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3819 +BUTTERFLYA(temp[8*1+i], temp[8*5+i])
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3820 +BUTTERFLYA(temp[8*2+i], temp[8*6+i])
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3821 +BUTTERFLYA(temp[8*3+i], temp[8*7+i]);
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3822 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3823
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3824 sum -= FFABS(temp[8*0] + temp[8*4]); // -mean
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3825
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3826 return sum;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3827 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3828
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3829 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
3830 MpegEncContext * const s= (MpegEncContext *)c;
11195
d464f498e19f Use LOCAL_ALIGNED macro for local arrays
mru
parents: 11193
diff changeset
3831 LOCAL_ALIGNED_16(DCTELEM, temp, [64]);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3832
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3833 assert(h==8);
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3834
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3835 s->dsp.diff_pixels(temp, src1, src2, stride);
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
3836 s->dsp.fdct(temp);
4988
689490842cf5 factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents: 4749
diff changeset
3837 return s->dsp.sum_abs_dctelem(temp);
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3838 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3839
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8496
diff changeset
3840 #if CONFIG_GPL
3010
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3841 #define DCT8_1D {\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3842 const int s07 = SRC(0) + SRC(7);\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3843 const int s16 = SRC(1) + SRC(6);\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3844 const int s25 = SRC(2) + SRC(5);\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3845 const int s34 = SRC(3) + SRC(4);\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3846 const int a0 = s07 + s34;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3847 const int a1 = s16 + s25;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3848 const int a2 = s07 - s34;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3849 const int a3 = s16 - s25;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3850 const int d07 = SRC(0) - SRC(7);\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3851 const int d16 = SRC(1) - SRC(6);\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3852 const int d25 = SRC(2) - SRC(5);\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3853 const int d34 = SRC(3) - SRC(4);\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3854 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
3855 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
3856 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
3857 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
3858 DST(0, a0 + a1 ) ;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3859 DST(1, a4 + (a7>>2)) ;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3860 DST(2, a2 + (a3>>1)) ;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3861 DST(3, a5 + (a6>>2)) ;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3862 DST(4, a0 - a1 ) ;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3863 DST(5, a6 - (a5>>2)) ;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3864 DST(6, (a2>>1) - a3 ) ;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3865 DST(7, (a4>>2) - a7 ) ;\
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3866 }
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3867
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3868 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
3869 MpegEncContext * const s= (MpegEncContext *)c;
5256
f540f39e1472 fix warning
mru
parents: 5255
diff changeset
3870 DCTELEM dct[8][8];
3010
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3871 int i;
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3872 int sum=0;
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3873
5256
f540f39e1472 fix warning
mru
parents: 5255
diff changeset
3874 s->dsp.diff_pixels(dct[0], src1, src2, stride);
3010
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3875
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3876 #define SRC(x) dct[i][x]
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3877 #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
3878 for( i = 0; i < 8; i++ )
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3879 DCT8_1D
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3880 #undef SRC
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3881 #undef DST
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3882
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3883 #define SRC(x) dct[x][i]
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3884 #define DST(x,v) sum += FFABS(v)
3010
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3885 for( i = 0; i < 8; i++ )
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3886 DCT8_1D
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3887 #undef SRC
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3888 #undef DST
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3889 return sum;
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3890 }
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3891 #endif
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
3892
2382
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3893 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
3894 MpegEncContext * const s= (MpegEncContext *)c;
11195
d464f498e19f Use LOCAL_ALIGNED macro for local arrays
mru
parents: 11193
diff changeset
3895 LOCAL_ALIGNED_16(DCTELEM, temp, [64]);
2382
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3896 int sum=0, i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3897
2382
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3898 assert(h==8);
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3899
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3900 s->dsp.diff_pixels(temp, src1, src2, stride);
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3901 s->dsp.fdct(temp);
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3902
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3903 for(i=0; i<64; i++)
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
3904 sum= FFMAX(sum, FFABS(temp[i]));
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3905
2382
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3906 return sum;
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3907 }
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
3908
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3909 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
3910 MpegEncContext * const s= (MpegEncContext *)c;
11195
d464f498e19f Use LOCAL_ALIGNED macro for local arrays
mru
parents: 11193
diff changeset
3911 LOCAL_ALIGNED_16(DCTELEM, temp, [64*2]);
11193
e5ebf3a17d9d Simplify some declarations of aligned arrays
mru
parents: 10965
diff changeset
3912 DCTELEM * const bak = temp+64;
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3913 int sum=0, i;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3914
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3915 assert(h==8);
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3916 s->mb_intra=0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3917
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3918 s->dsp.diff_pixels(temp, src1, src2, stride);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3919
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3920 memcpy(bak, temp, 64*sizeof(DCTELEM));
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3921
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3922 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
3923 s->dct_unquantize_inter(s, temp, 0, s->qscale);
6001
f4859c13426b add ff_ prefix to all simple_idct symbols
aurel
parents: 5899
diff changeset
3924 ff_simple_idct(temp); //FIXME
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3925
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3926 for(i=0; i<64; i++)
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3927 sum+= (temp[i]-bak[i])*(temp[i]-bak[i]);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3928
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3929 return sum;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3930 }
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
3931
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3932 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
3933 MpegEncContext * const s= (MpegEncContext *)c;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
3934 const uint8_t *scantable= s->intra_scantable.permutated;
11195
d464f498e19f Use LOCAL_ALIGNED macro for local arrays
mru
parents: 11193
diff changeset
3935 LOCAL_ALIGNED_16(DCTELEM, temp, [64]);
d464f498e19f Use LOCAL_ALIGNED macro for local arrays
mru
parents: 11193
diff changeset
3936 LOCAL_ALIGNED_16(uint8_t, lsrc1, [64]);
d464f498e19f Use LOCAL_ALIGNED macro for local arrays
mru
parents: 11193
diff changeset
3937 LOCAL_ALIGNED_16(uint8_t, lsrc2, [64]);
6719
e88e719b5e77 Typo: distoration -> distortion.
ramiro
parents: 6600
diff changeset
3938 int i, last, run, bits, level, distortion, start_i;
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3939 const int esc_length= s->ac_esc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3940 uint8_t * length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3941 uint8_t * last_length;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3942
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3943 assert(h==8);
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
3944
10068
770294452eaf Get rid of large stack array in rd8x8_c()
mru
parents: 9995
diff changeset
3945 copy_block8(lsrc1, src1, 8, stride, 8);
770294452eaf Get rid of large stack array in rd8x8_c()
mru
parents: 9995
diff changeset
3946 copy_block8(lsrc2, src2, 8, stride, 8);
770294452eaf Get rid of large stack array in rd8x8_c()
mru
parents: 9995
diff changeset
3947
770294452eaf Get rid of large stack array in rd8x8_c()
mru
parents: 9995
diff changeset
3948 s->dsp.diff_pixels(temp, lsrc1, lsrc2, 8);
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3949
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3950 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
3951
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3952 bits=0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3953
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3954 if (s->mb_intra) {
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3955 start_i = 1;
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3956 length = s->intra_ac_vlc_length;
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3957 last_length= s->intra_ac_vlc_last_length;
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3958 bits+= s->luma_dc_vlc_length[temp[0] + 256]; //FIXME chroma
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3959 } else {
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3960 start_i = 0;
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3961 length = s->inter_ac_vlc_length;
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3962 last_length= s->inter_ac_vlc_last_length;
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3963 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3964
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3965 if(last>=start_i){
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3966 run=0;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3967 for(i=start_i; i<last; i++){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3968 int j= scantable[i];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3969 level= temp[j];
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3970
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3971 if(level){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3972 level+=64;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3973 if((level&(~127)) == 0){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3974 bits+= length[UNI_AC_ENC_INDEX(run, level)];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3975 }else
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3976 bits+= esc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3977 run=0;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3978 }else
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3979 run++;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3980 }
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3981 i= scantable[last];
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3982
1011
michaelni
parents: 1008
diff changeset
3983 level= temp[i] + 64;
michaelni
parents: 1008
diff changeset
3984
michaelni
parents: 1008
diff changeset
3985 assert(level - 64);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3986
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3987 if((level&(~127)) == 0){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3988 bits+= last_length[UNI_AC_ENC_INDEX(run, level)];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3989 }else
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3990 bits+= esc_length;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
3991
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3992 }
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3993
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
3994 if(last>=0){
1689
1a2db2073848 split intra / inter dequantization
michael
parents: 1645
diff changeset
3995 if(s->mb_intra)
1a2db2073848 split intra / inter dequantization
michael
parents: 1645
diff changeset
3996 s->dct_unquantize_intra(s, temp, 0, s->qscale);
1a2db2073848 split intra / inter dequantization
michael
parents: 1645
diff changeset
3997 else
1a2db2073848 split intra / inter dequantization
michael
parents: 1645
diff changeset
3998 s->dct_unquantize_inter(s, temp, 0, s->qscale);
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
3999 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4000
10068
770294452eaf Get rid of large stack array in rd8x8_c()
mru
parents: 9995
diff changeset
4001 s->dsp.idct_add(lsrc2, 8, temp);
770294452eaf Get rid of large stack array in rd8x8_c()
mru
parents: 9995
diff changeset
4002
770294452eaf Get rid of large stack array in rd8x8_c()
mru
parents: 9995
diff changeset
4003 distortion= s->dsp.sse[1](NULL, lsrc2, lsrc1, 8, 8);
6719
e88e719b5e77 Typo: distoration -> distortion.
ramiro
parents: 6600
diff changeset
4004
e88e719b5e77 Typo: distoration -> distortion.
ramiro
parents: 6600
diff changeset
4005 return distortion + ((bits*s->qscale*s->qscale*109 + 64)>>7);
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4006 }
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4007
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4008 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
4009 MpegEncContext * const s= (MpegEncContext *)c;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1016
diff changeset
4010 const uint8_t *scantable= s->intra_scantable.permutated;
11195
d464f498e19f Use LOCAL_ALIGNED macro for local arrays
mru
parents: 11193
diff changeset
4011 LOCAL_ALIGNED_16(DCTELEM, temp, [64]);
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4012 int i, last, run, bits, level, start_i;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4013 const int esc_length= s->ac_esc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4014 uint8_t * length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4015 uint8_t * last_length;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4016
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4017 assert(h==8);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4018
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
4019 s->dsp.diff_pixels(temp, src1, src2, stride);
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4020
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
4021 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
4022
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
4023 bits=0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4024
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4025 if (s->mb_intra) {
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4026 start_i = 1;
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4027 length = s->intra_ac_vlc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4028 last_length= s->intra_ac_vlc_last_length;
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
4029 bits+= s->luma_dc_vlc_length[temp[0] + 256]; //FIXME chroma
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4030 } else {
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4031 start_i = 0;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4032 length = s->inter_ac_vlc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4033 last_length= s->inter_ac_vlc_last_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4034 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4035
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
4036 if(last>=start_i){
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4037 run=0;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4038 for(i=start_i; i<last; i++){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4039 int j= scantable[i];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4040 level= temp[j];
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4041
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4042 if(level){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4043 level+=64;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4044 if((level&(~127)) == 0){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4045 bits+= length[UNI_AC_ENC_INDEX(run, level)];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4046 }else
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4047 bits+= esc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4048 run=0;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4049 }else
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4050 run++;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4051 }
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4052 i= scantable[last];
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4053
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
4054 level= temp[i] + 64;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4055
1013
5d4c95f323d0 finetuneing thresholds/factors
michaelni
parents: 1012
diff changeset
4056 assert(level - 64);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4057
1007
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4058 if((level&(~127)) == 0){
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4059 bits+= last_length[UNI_AC_ENC_INDEX(run, level)];
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4060 }else
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4061 bits+= esc_length;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4062 }
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4063
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4064 return bits;
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4065 }
b2cf2a1d9a51 more compare functions (rd & bit)
michaelni
parents: 997
diff changeset
4066
8978
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4067 #define VSAD_INTRA(size) \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4068 static int vsad_intra##size##_c(/*MpegEncContext*/ void *c, uint8_t *s, uint8_t *dummy, int stride, int h){ \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4069 int score=0; \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4070 int x,y; \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4071 \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4072 for(y=1; y<h; y++){ \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4073 for(x=0; x<size; x+=4){ \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4074 score+= FFABS(s[x ] - s[x +stride]) + FFABS(s[x+1] - s[x+1+stride]) \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4075 +FFABS(s[x+2] - s[x+2+stride]) + FFABS(s[x+3] - s[x+3+stride]); \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4076 } \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4077 s+= stride; \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4078 } \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4079 \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4080 return score; \
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4081 }
8978
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4082 VSAD_INTRA(8)
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4083 VSAD_INTRA(16)
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4084
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4085 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
4086 int score=0;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4087 int x,y;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4088
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4089 for(y=1; y<h; y++){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4090 for(x=0; x<16; x++){
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3947
diff changeset
4091 score+= FFABS(s1[x ] - s2[x ] - s1[x +stride] + s2[x +stride]);
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4092 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4093 s1+= stride;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4094 s2+= stride;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4095 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4096
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4097 return score;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4098 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4099
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4100 #define SQ(a) ((a)*(a))
8978
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4101 #define VSSE_INTRA(size) \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4102 static int vsse_intra##size##_c(/*MpegEncContext*/ void *c, uint8_t *s, uint8_t *dummy, int stride, int h){ \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4103 int score=0; \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4104 int x,y; \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4105 \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4106 for(y=1; y<h; y++){ \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4107 for(x=0; x<size; x+=4){ \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4108 score+= SQ(s[x ] - s[x +stride]) + SQ(s[x+1] - s[x+1+stride]) \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4109 +SQ(s[x+2] - s[x+2+stride]) + SQ(s[x+3] - s[x+3+stride]); \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4110 } \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4111 s+= stride; \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4112 } \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4113 \
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4114 return score; \
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4115 }
8978
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4116 VSSE_INTRA(8)
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4117 VSSE_INTRA(16)
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4118
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4119 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
4120 int score=0;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4121 int x,y;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4122
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4123 for(y=1; y<h; y++){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4124 for(x=0; x<16; x++){
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4125 score+= SQ(s1[x ] - s2[x ] - s1[x +stride] + s2[x +stride]);
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4126 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4127 s1+= stride;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4128 s2+= stride;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4129 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4130
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4131 return score;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4132 }
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4133
5255
669a97223dc7 make arguments to ssd_int8_vs_int16() const
mru
parents: 5214
diff changeset
4134 static int ssd_int8_vs_int16_c(const int8_t *pix1, const int16_t *pix2,
669a97223dc7 make arguments to ssd_int8_vs_int16() const
mru
parents: 5214
diff changeset
4135 int size){
4749
7011f597e473 mmx 16-bit ssd. 2.3x faster svq1 encoding.
lorenm
parents: 4598
diff changeset
4136 int score=0;
7011f597e473 mmx 16-bit ssd. 2.3x faster svq1 encoding.
lorenm
parents: 4598
diff changeset
4137 int i;
7011f597e473 mmx 16-bit ssd. 2.3x faster svq1 encoding.
lorenm
parents: 4598
diff changeset
4138 for(i=0; i<size; i++)
7011f597e473 mmx 16-bit ssd. 2.3x faster svq1 encoding.
lorenm
parents: 4598
diff changeset
4139 score += (pix1[i]-pix2[i])*(pix1[i]-pix2[i]);
7011f597e473 mmx 16-bit ssd. 2.3x faster svq1 encoding.
lorenm
parents: 4598
diff changeset
4140 return score;
7011f597e473 mmx 16-bit ssd. 2.3x faster svq1 encoding.
lorenm
parents: 4598
diff changeset
4141 }
7011f597e473 mmx 16-bit ssd. 2.3x faster svq1 encoding.
lorenm
parents: 4598
diff changeset
4142
6056
558c1fd0ee72 Fix typo in macro name: WARPER8_16_SQ --> WRAPPER8_16_SQ.
diego
parents: 6054
diff changeset
4143 WRAPPER8_16_SQ(hadamard8_diff8x8_c, hadamard8_diff16_c)
558c1fd0ee72 Fix typo in macro name: WARPER8_16_SQ --> WRAPPER8_16_SQ.
diego
parents: 6054
diff changeset
4144 WRAPPER8_16_SQ(hadamard8_intra8x8_c, hadamard8_intra16_c)
558c1fd0ee72 Fix typo in macro name: WARPER8_16_SQ --> WRAPPER8_16_SQ.
diego
parents: 6054
diff changeset
4145 WRAPPER8_16_SQ(dct_sad8x8_c, dct_sad16_c)
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8496
diff changeset
4146 #if CONFIG_GPL
6056
558c1fd0ee72 Fix typo in macro name: WARPER8_16_SQ --> WRAPPER8_16_SQ.
diego
parents: 6054
diff changeset
4147 WRAPPER8_16_SQ(dct264_sad8x8_c, dct264_sad16_c)
3013
b6a90fabaa76 ugly, but necessary
melanson
parents: 3010
diff changeset
4148 #endif
6056
558c1fd0ee72 Fix typo in macro name: WARPER8_16_SQ --> WRAPPER8_16_SQ.
diego
parents: 6054
diff changeset
4149 WRAPPER8_16_SQ(dct_max8x8_c, dct_max16_c)
558c1fd0ee72 Fix typo in macro name: WARPER8_16_SQ --> WRAPPER8_16_SQ.
diego
parents: 6054
diff changeset
4150 WRAPPER8_16_SQ(quant_psnr8x8_c, quant_psnr16_c)
558c1fd0ee72 Fix typo in macro name: WARPER8_16_SQ --> WRAPPER8_16_SQ.
diego
parents: 6054
diff changeset
4151 WRAPPER8_16_SQ(rd8x8_c, rd16_c)
558c1fd0ee72 Fix typo in macro name: WARPER8_16_SQ --> WRAPPER8_16_SQ.
diego
parents: 6054
diff changeset
4152 WRAPPER8_16_SQ(bit8x8_c, bit16_c)
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
4153
3568
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4154 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
4155 int i;
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4156 for(i=0; i<len; i++)
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4157 dst[i] *= src[i];
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4158 }
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4159
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4160 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
4161 int i;
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4162 src1 += len-1;
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4163 for(i=0; i<len; i++)
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4164 dst[i] = src0[i] * src1[-i];
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4165 }
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4166
10300
4d1b9ca628fc Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents: 10219
diff changeset
4167 static void vector_fmul_add_c(float *dst, const float *src0, const float *src1, const float *src2, int len){
3568
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4168 int i;
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4169 for(i=0; i<len; i++)
10300
4d1b9ca628fc Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents: 10219
diff changeset
4170 dst[i] = src0[i] * src1[i] + src2[i];
3568
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4171 }
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4172
7261
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4173 void ff_vector_fmul_window_c(float *dst, const float *src0, const float *src1, const float *win, float add_bias, int len){
7263
fc843d00867c exploit mdct symmetry
lorenm
parents: 7261
diff changeset
4174 int i,j;
fc843d00867c exploit mdct symmetry
lorenm
parents: 7261
diff changeset
4175 dst += len;
fc843d00867c exploit mdct symmetry
lorenm
parents: 7261
diff changeset
4176 win += len;
fc843d00867c exploit mdct symmetry
lorenm
parents: 7261
diff changeset
4177 src0+= len;
fc843d00867c exploit mdct symmetry
lorenm
parents: 7261
diff changeset
4178 for(i=-len, j=len-1; i<0; i++, j--) {
fc843d00867c exploit mdct symmetry
lorenm
parents: 7261
diff changeset
4179 float s0 = src0[i];
fc843d00867c exploit mdct symmetry
lorenm
parents: 7261
diff changeset
4180 float s1 = src1[j];
fc843d00867c exploit mdct symmetry
lorenm
parents: 7261
diff changeset
4181 float wi = win[i];
fc843d00867c exploit mdct symmetry
lorenm
parents: 7261
diff changeset
4182 float wj = win[j];
fc843d00867c exploit mdct symmetry
lorenm
parents: 7261
diff changeset
4183 dst[i] = s0*wj - s1*wi + add_bias;
fc843d00867c exploit mdct symmetry
lorenm
parents: 7261
diff changeset
4184 dst[j] = s0*wi + s1*wj + add_bias;
fc843d00867c exploit mdct symmetry
lorenm
parents: 7261
diff changeset
4185 }
7261
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4186 }
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4187
10219
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4188 static void vector_fmul_scalar_c(float *dst, const float *src, float mul,
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4189 int len)
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4190 {
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4191 int i;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4192 for (i = 0; i < len; i++)
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4193 dst[i] = src[i] * mul;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4194 }
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4195
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4196 static void vector_fmul_sv_scalar_2_c(float *dst, const float *src,
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4197 const float **sv, float mul, int len)
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4198 {
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4199 int i;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4200 for (i = 0; i < len; i += 2, sv++) {
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4201 dst[i ] = src[i ] * sv[0][0] * mul;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4202 dst[i+1] = src[i+1] * sv[0][1] * mul;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4203 }
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4204 }
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4205
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4206 static void vector_fmul_sv_scalar_4_c(float *dst, const float *src,
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4207 const float **sv, float mul, int len)
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4208 {
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4209 int i;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4210 for (i = 0; i < len; i += 4, sv++) {
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4211 dst[i ] = src[i ] * sv[0][0] * mul;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4212 dst[i+1] = src[i+1] * sv[0][1] * mul;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4213 dst[i+2] = src[i+2] * sv[0][2] * mul;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4214 dst[i+3] = src[i+3] * sv[0][3] * mul;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4215 }
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4216 }
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4217
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4218 static void sv_fmul_scalar_2_c(float *dst, const float **sv, float mul,
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4219 int len)
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4220 {
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4221 int i;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4222 for (i = 0; i < len; i += 2, sv++) {
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4223 dst[i ] = sv[0][0] * mul;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4224 dst[i+1] = sv[0][1] * mul;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4225 }
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4226 }
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4227
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4228 static void sv_fmul_scalar_4_c(float *dst, const float **sv, float mul,
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4229 int len)
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4230 {
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4231 int i;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4232 for (i = 0; i < len; i += 4, sv++) {
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4233 dst[i ] = sv[0][0] * mul;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4234 dst[i+1] = sv[0][1] * mul;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4235 dst[i+2] = sv[0][2] * mul;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4236 dst[i+3] = sv[0][3] * mul;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4237 }
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4238 }
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4239
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4240 static void butterflies_float_c(float *restrict v1, float *restrict v2,
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4241 int len)
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4242 {
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4243 int i;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4244 for (i = 0; i < len; i++) {
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4245 float t = v1[i] - v2[i];
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4246 v1[i] += v2[i];
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4247 v2[i] = t;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4248 }
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4249 }
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4250
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4251 static float scalarproduct_float_c(const float *v1, const float *v2, int len)
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4252 {
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4253 float p = 0.0;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4254 int i;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4255
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4256 for (i = 0; i < len; i++)
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4257 p += v1[i] * v2[i];
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4258
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4259 return p;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4260 }
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4261
7564
7cf793954871 simd int->float
lorenm
parents: 7563
diff changeset
4262 static void int32_to_float_fmul_scalar_c(float *dst, const int *src, float mul, int len){
7cf793954871 simd int->float
lorenm
parents: 7563
diff changeset
4263 int i;
7cf793954871 simd int->float
lorenm
parents: 7563
diff changeset
4264 for(i=0; i<len; i++)
7cf793954871 simd int->float
lorenm
parents: 7563
diff changeset
4265 dst[i] = src[i] * mul;
7cf793954871 simd int->float
lorenm
parents: 7563
diff changeset
4266 }
7cf793954871 simd int->float
lorenm
parents: 7563
diff changeset
4267
10104
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4268 static inline uint32_t clipf_c_one(uint32_t a, uint32_t mini,
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4269 uint32_t maxi, uint32_t maxisign)
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4270 {
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4271
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4272 if(a > mini) return mini;
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4273 else if((a^(1<<31)) > maxisign) return maxi;
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4274 else return a;
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4275 }
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4276
10105
7775f6627612 Mark parameter src of vector_clipf() as const
vitor
parents: 10104
diff changeset
4277 static void vector_clipf_c_opposite_sign(float *dst, const float *src, float *min, float *max, int len){
10104
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4278 int i;
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4279 uint32_t mini = *(uint32_t*)min;
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4280 uint32_t maxi = *(uint32_t*)max;
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4281 uint32_t maxisign = maxi ^ (1<<31);
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4282 uint32_t *dsti = (uint32_t*)dst;
10105
7775f6627612 Mark parameter src of vector_clipf() as const
vitor
parents: 10104
diff changeset
4283 const uint32_t *srci = (const uint32_t*)src;
10104
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4284 for(i=0; i<len; i+=8) {
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4285 dsti[i + 0] = clipf_c_one(srci[i + 0], mini, maxi, maxisign);
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4286 dsti[i + 1] = clipf_c_one(srci[i + 1], mini, maxi, maxisign);
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4287 dsti[i + 2] = clipf_c_one(srci[i + 2], mini, maxi, maxisign);
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4288 dsti[i + 3] = clipf_c_one(srci[i + 3], mini, maxi, maxisign);
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4289 dsti[i + 4] = clipf_c_one(srci[i + 4], mini, maxi, maxisign);
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4290 dsti[i + 5] = clipf_c_one(srci[i + 5], mini, maxi, maxisign);
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4291 dsti[i + 6] = clipf_c_one(srci[i + 6], mini, maxi, maxisign);
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4292 dsti[i + 7] = clipf_c_one(srci[i + 7], mini, maxi, maxisign);
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4293 }
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4294 }
10105
7775f6627612 Mark parameter src of vector_clipf() as const
vitor
parents: 10104
diff changeset
4295 static void vector_clipf_c(float *dst, const float *src, float min, float max, int len){
10104
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4296 int i;
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4297 if(min < 0 && max > 0) {
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4298 vector_clipf_c_opposite_sign(dst, src, &min, &max, len);
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4299 } else {
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4300 for(i=0; i < len; i+=8) {
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4301 dst[i ] = av_clipf(src[i ], min, max);
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4302 dst[i + 1] = av_clipf(src[i + 1], min, max);
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4303 dst[i + 2] = av_clipf(src[i + 2], min, max);
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4304 dst[i + 3] = av_clipf(src[i + 3], min, max);
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4305 dst[i + 4] = av_clipf(src[i + 4], min, max);
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4306 dst[i + 5] = av_clipf(src[i + 5], min, max);
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4307 dst[i + 6] = av_clipf(src[i + 6], min, max);
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4308 dst[i + 7] = av_clipf(src[i + 7], min, max);
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4309 }
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4310 }
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4311 }
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4312
7261
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4313 static av_always_inline int float_to_int16_one(const float *src){
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4314 int_fast32_t tmp = *(const int32_t*)src;
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4315 if(tmp & 0xf0000){
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4316 tmp = (0x43c0ffff - tmp)>>31;
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4317 // is this faster on some gcc/cpu combinations?
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4318 // if(tmp > 0x43c0ffff) tmp = 0xFFFF;
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4319 // else tmp = 0;
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4320 }
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4321 return tmp - 0x8000;
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4322 }
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4323
7218
7f3d6509628b Fix x86-64
michael
parents: 7203
diff changeset
4324 void ff_float_to_int16_c(int16_t *dst, const float *src, long len){
3568
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4325 int i;
7261
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4326 for(i=0; i<len; i++)
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4327 dst[i] = float_to_int16_one(src+i);
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4328 }
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4329
7286
e267f2519248 float_to_int16_interleave: change src to an array of pointers instead of assuming it's contiguous.
lorenm
parents: 7263
diff changeset
4330 void ff_float_to_int16_interleave_c(int16_t *dst, const float **src, long len, int channels){
7261
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4331 int i,j,c;
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4332 if(channels==2){
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4333 for(i=0; i<len; i++){
7286
e267f2519248 float_to_int16_interleave: change src to an array of pointers instead of assuming it's contiguous.
lorenm
parents: 7263
diff changeset
4334 dst[2*i] = float_to_int16_one(src[0]+i);
e267f2519248 float_to_int16_interleave: change src to an array of pointers instead of assuming it's contiguous.
lorenm
parents: 7263
diff changeset
4335 dst[2*i+1] = float_to_int16_one(src[1]+i);
3568
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4336 }
7261
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4337 }else{
7286
e267f2519248 float_to_int16_interleave: change src to an array of pointers instead of assuming it's contiguous.
lorenm
parents: 7263
diff changeset
4338 for(c=0; c<channels; c++)
7261
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4339 for(i=0, j=c; i<len; i++, j+=channels)
7286
e267f2519248 float_to_int16_interleave: change src to an array of pointers instead of assuming it's contiguous.
lorenm
parents: 7263
diff changeset
4340 dst[j] = float_to_int16_one(src[c]+i);
3568
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4341 }
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4342 }
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4343
7203
87b1dfb5a98d Add several vector functions used by Monkey's Audio decoder to dsputil
kostya
parents: 6719
diff changeset
4344 static int32_t scalarproduct_int16_c(int16_t * v1, int16_t * v2, int order, int shift)
87b1dfb5a98d Add several vector functions used by Monkey's Audio decoder to dsputil
kostya
parents: 6719
diff changeset
4345 {
87b1dfb5a98d Add several vector functions used by Monkey's Audio decoder to dsputil
kostya
parents: 6719
diff changeset
4346 int res = 0;
87b1dfb5a98d Add several vector functions used by Monkey's Audio decoder to dsputil
kostya
parents: 6719
diff changeset
4347
87b1dfb5a98d Add several vector functions used by Monkey's Audio decoder to dsputil
kostya
parents: 6719
diff changeset
4348 while (order--)
87b1dfb5a98d Add several vector functions used by Monkey's Audio decoder to dsputil
kostya
parents: 6719
diff changeset
4349 res += (*v1++ * *v2++) >> shift;
87b1dfb5a98d Add several vector functions used by Monkey's Audio decoder to dsputil
kostya
parents: 6719
diff changeset
4350
87b1dfb5a98d Add several vector functions used by Monkey's Audio decoder to dsputil
kostya
parents: 6719
diff changeset
4351 return res;
87b1dfb5a98d Add several vector functions used by Monkey's Audio decoder to dsputil
kostya
parents: 6719
diff changeset
4352 }
87b1dfb5a98d Add several vector functions used by Monkey's Audio decoder to dsputil
kostya
parents: 6719
diff changeset
4353
10644
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10431
diff changeset
4354 static int32_t scalarproduct_and_madd_int16_c(int16_t *v1, int16_t *v2, int16_t *v3, int order, int mul)
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10431
diff changeset
4355 {
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10431
diff changeset
4356 int res = 0;
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10431
diff changeset
4357 while (order--) {
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10431
diff changeset
4358 res += *v1 * *v2++;
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10431
diff changeset
4359 *v1++ += mul * *v3++;
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10431
diff changeset
4360 }
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10431
diff changeset
4361 return res;
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10431
diff changeset
4362 }
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10431
diff changeset
4363
5887
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4364 #define W0 2048
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4365 #define W1 2841 /* 2048*sqrt (2)*cos (1*pi/16) */
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4366 #define W2 2676 /* 2048*sqrt (2)*cos (2*pi/16) */
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4367 #define W3 2408 /* 2048*sqrt (2)*cos (3*pi/16) */
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4368 #define W4 2048 /* 2048*sqrt (2)*cos (4*pi/16) */
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4369 #define W5 1609 /* 2048*sqrt (2)*cos (5*pi/16) */
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4370 #define W6 1108 /* 2048*sqrt (2)*cos (6*pi/16) */
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4371 #define W7 565 /* 2048*sqrt (2)*cos (7*pi/16) */
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4372
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4373 static void wmv2_idct_row(short * b)
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4374 {
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4375 int s1,s2;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4376 int a0,a1,a2,a3,a4,a5,a6,a7;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4377 /*step 1*/
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4378 a1 = W1*b[1]+W7*b[7];
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4379 a7 = W7*b[1]-W1*b[7];
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4380 a5 = W5*b[5]+W3*b[3];
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4381 a3 = W3*b[5]-W5*b[3];
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4382 a2 = W2*b[2]+W6*b[6];
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4383 a6 = W6*b[2]-W2*b[6];
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4384 a0 = W0*b[0]+W0*b[4];
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4385 a4 = W0*b[0]-W0*b[4];
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4386 /*step 2*/
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4387 s1 = (181*(a1-a5+a7-a3)+128)>>8;//1,3,5,7,
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4388 s2 = (181*(a1-a5-a7+a3)+128)>>8;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4389 /*step 3*/
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4390 b[0] = (a0+a2+a1+a5 + (1<<7))>>8;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4391 b[1] = (a4+a6 +s1 + (1<<7))>>8;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4392 b[2] = (a4-a6 +s2 + (1<<7))>>8;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4393 b[3] = (a0-a2+a7+a3 + (1<<7))>>8;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4394 b[4] = (a0-a2-a7-a3 + (1<<7))>>8;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4395 b[5] = (a4-a6 -s2 + (1<<7))>>8;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4396 b[6] = (a4+a6 -s1 + (1<<7))>>8;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4397 b[7] = (a0+a2-a1-a5 + (1<<7))>>8;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4398 }
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4399 static void wmv2_idct_col(short * b)
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4400 {
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4401 int s1,s2;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4402 int a0,a1,a2,a3,a4,a5,a6,a7;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4403 /*step 1, with extended precision*/
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4404 a1 = (W1*b[8*1]+W7*b[8*7] + 4)>>3;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4405 a7 = (W7*b[8*1]-W1*b[8*7] + 4)>>3;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4406 a5 = (W5*b[8*5]+W3*b[8*3] + 4)>>3;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4407 a3 = (W3*b[8*5]-W5*b[8*3] + 4)>>3;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4408 a2 = (W2*b[8*2]+W6*b[8*6] + 4)>>3;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4409 a6 = (W6*b[8*2]-W2*b[8*6] + 4)>>3;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4410 a0 = (W0*b[8*0]+W0*b[8*4] )>>3;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4411 a4 = (W0*b[8*0]-W0*b[8*4] )>>3;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4412 /*step 2*/
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4413 s1 = (181*(a1-a5+a7-a3)+128)>>8;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4414 s2 = (181*(a1-a5-a7+a3)+128)>>8;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4415 /*step 3*/
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4416 b[8*0] = (a0+a2+a1+a5 + (1<<13))>>14;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4417 b[8*1] = (a4+a6 +s1 + (1<<13))>>14;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4418 b[8*2] = (a4-a6 +s2 + (1<<13))>>14;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4419 b[8*3] = (a0-a2+a7+a3 + (1<<13))>>14;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4420
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4421 b[8*4] = (a0-a2-a7-a3 + (1<<13))>>14;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4422 b[8*5] = (a4-a6 -s2 + (1<<13))>>14;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4423 b[8*6] = (a4+a6 -s1 + (1<<13))>>14;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4424 b[8*7] = (a0+a2-a1-a5 + (1<<13))>>14;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4425 }
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4426 void ff_wmv2_idct_c(short * block){
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4427 int i;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4428
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4429 for(i=0;i<64;i+=8){
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4430 wmv2_idct_row(block+i);
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4431 }
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4432 for(i=0;i<8;i++){
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4433 wmv2_idct_col(block+i);
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4434 }
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4435 }
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4436 /* XXX: those functions should be suppressed ASAP when all IDCTs are
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4437 converted */
5887
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4438 static void ff_wmv2_idct_put_c(uint8_t *dest, int line_size, DCTELEM *block)
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4439 {
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4440 ff_wmv2_idct_c(block);
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4441 put_pixels_clamped_c(block, dest, line_size);
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4442 }
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4443 static void ff_wmv2_idct_add_c(uint8_t *dest, int line_size, DCTELEM *block)
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4444 {
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4445 ff_wmv2_idct_c(block);
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4446 add_pixels_clamped_c(block, dest, line_size);
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4447 }
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4448 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
4449 {
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4450 j_rev_dct (block);
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4451 put_pixels_clamped_c(block, dest, line_size);
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4452 }
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4453 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
4454 {
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4455 j_rev_dct (block);
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4456 add_pixels_clamped_c(block, dest, line_size);
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4457 }
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4458
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4459 static void ff_jref_idct4_put(uint8_t *dest, int line_size, DCTELEM *block)
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4460 {
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4461 j_rev_dct4 (block);
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4462 put_pixels_clamped4_c(block, dest, line_size);
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4463 }
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4464 static void ff_jref_idct4_add(uint8_t *dest, int line_size, DCTELEM *block)
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4465 {
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4466 j_rev_dct4 (block);
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4467 add_pixels_clamped4_c(block, dest, line_size);
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4468 }
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4469
2257
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
4470 static void ff_jref_idct2_put(uint8_t *dest, int line_size, DCTELEM *block)
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
4471 {
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
4472 j_rev_dct2 (block);
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
4473 put_pixels_clamped2_c(block, dest, line_size);
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
4474 }
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
4475 static void ff_jref_idct2_add(uint8_t *dest, int line_size, DCTELEM *block)
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
4476 {
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
4477 j_rev_dct2 (block);
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
4478 add_pixels_clamped2_c(block, dest, line_size);
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
4479 }
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
4480
2259
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
4481 static void ff_jref_idct1_put(uint8_t *dest, int line_size, DCTELEM *block)
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
4482 {
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
4483 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;
2259
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
4484
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
4485 dest[0] = cm[(block[0] + 4)>>3];
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
4486 }
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
4487 static void ff_jref_idct1_add(uint8_t *dest, int line_size, DCTELEM *block)
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
4488 {
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
4489 uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;
2259
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
4490
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
4491 dest[0] = cm[dest[0] + ((block[0] + 4)>>3)];
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
4492 }
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
4493
5143
151f6f3a507b Correct signature of just_return().
takis
parents: 5140
diff changeset
4494 static void just_return(void *mem av_unused, int stride av_unused, int h av_unused) { return; }
3215
06f98047ff26 prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents: 3199
diff changeset
4495
1201
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
4496 /* init static data */
10867
bf309c7ce615 More av_cold for infrequently called functions.
zuxy
parents: 10749
diff changeset
4497 av_cold void dsputil_static_init(void)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
4498 {
751
cbe316f082bc warning fixes
michaelni
parents: 706
diff changeset
4499 int i;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
4500
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
4501 for(i=0;i<256;i++) ff_cropTbl[i + MAX_NEG_CROP] = i;
1201
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
4502 for(i=0;i<MAX_NEG_CROP;i++) {
4176
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
4503 ff_cropTbl[i] = 0;
23da44e8fd05 rename cropTbl -> ff_cropTbl
mru
parents: 4174
diff changeset
4504 ff_cropTbl[i + MAX_NEG_CROP + 256] = 255;
1201
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
4505 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4506
1201
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
4507 for(i=0;i<512;i++) {
4179
46d38e3d7038 rename squareTbl -> ff_squareTbl
mru
parents: 4176
diff changeset
4508 ff_squareTbl[i] = (i - 256) * (i - 256);
1201
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
4509 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4510
4197
bbe0bc387a19 revert bad checkin
mru
parents: 4196
diff changeset
4511 for(i=0; i<64; i++) inv_zigzag_direct16[ff_zigzag_direct[i]]= i+1;
1201
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
4512 }
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
4513
4281
de525a2b41db ff_check_alignment to warn the user about a missaligned stack
michael
parents: 4240
diff changeset
4514 int ff_check_alignment(void){
de525a2b41db ff_check_alignment to warn the user about a missaligned stack
michael
parents: 4240
diff changeset
4515 static int did_fail=0;
11369
98970e51365a Remove DECLARE_ALIGNED_{8,16} macros
mru
parents: 11231
diff changeset
4516 DECLARE_ALIGNED(16, int, aligned);
4281
de525a2b41db ff_check_alignment to warn the user about a missaligned stack
michael
parents: 4240
diff changeset
4517
9259
a031926f7d6b Use intptr_t when casting pointers to int.
ramiro
parents: 8978
diff changeset
4518 if((intptr_t)&aligned & 15){
4281
de525a2b41db ff_check_alignment to warn the user about a missaligned stack
michael
parents: 4240
diff changeset
4519 if(!did_fail){
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8496
diff changeset
4520 #if HAVE_MMX || HAVE_ALTIVEC
4281
de525a2b41db ff_check_alignment to warn the user about a missaligned stack
michael
parents: 4240
diff changeset
4521 av_log(NULL, AV_LOG_ERROR,
4292
a26b5096dbb0 improved stack misalignment warning
mru
parents: 4282
diff changeset
4522 "Compiler did not align stack variables. Libavcodec has been miscompiled\n"
a26b5096dbb0 improved stack misalignment warning
mru
parents: 4282
diff changeset
4523 "and may be very slow or crash. This is not a bug in libavcodec,\n"
5542
b0a566346fb1 Add attribute that forces alignment of stack to functions that need it.
ramiro
parents: 5520
diff changeset
4524 "but in the compiler. You may try recompiling using gcc >= 4.2.\n"
b0a566346fb1 Add attribute that forces alignment of stack to functions that need it.
ramiro
parents: 5520
diff changeset
4525 "Do not report crashes to FFmpeg developers.\n");
4281
de525a2b41db ff_check_alignment to warn the user about a missaligned stack
michael
parents: 4240
diff changeset
4526 #endif
de525a2b41db ff_check_alignment to warn the user about a missaligned stack
michael
parents: 4240
diff changeset
4527 did_fail=1;
de525a2b41db ff_check_alignment to warn the user about a missaligned stack
michael
parents: 4240
diff changeset
4528 }
de525a2b41db ff_check_alignment to warn the user about a missaligned stack
michael
parents: 4240
diff changeset
4529 return -1;
de525a2b41db ff_check_alignment to warn the user about a missaligned stack
michael
parents: 4240
diff changeset
4530 }
de525a2b41db ff_check_alignment to warn the user about a missaligned stack
michael
parents: 4240
diff changeset
4531 return 0;
de525a2b41db ff_check_alignment to warn the user about a missaligned stack
michael
parents: 4240
diff changeset
4532 }
861
243cc33da3eb * init for inv_zigzag_direct16 moved to init block
kabi
parents: 860
diff changeset
4533
10867
bf309c7ce615 More av_cold for infrequently called functions.
zuxy
parents: 10749
diff changeset
4534 av_cold void dsputil_init(DSPContext* c, AVCodecContext *avctx)
1201
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
4535 {
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
4536 int i;
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
4537
4281
de525a2b41db ff_check_alignment to warn the user about a missaligned stack
michael
parents: 4240
diff changeset
4538 ff_check_alignment();
de525a2b41db ff_check_alignment to warn the user about a missaligned stack
michael
parents: 4240
diff changeset
4539
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8496
diff changeset
4540 #if CONFIG_ENCODERS
1567
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
4541 if(avctx->dct_algo==FF_DCT_FASTINT) {
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4542 c->fdct = fdct_ifast;
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
4543 c->fdct248 = fdct_ifast248;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4544 }
1567
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
4545 else if(avctx->dct_algo==FF_DCT_FAAN) {
1557
5d53c03186a1 floating point AAN DCT
michael
parents: 1527
diff changeset
4546 c->fdct = ff_faandct;
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
4547 c->fdct248 = ff_faandct248;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4548 }
1567
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
4549 else {
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4550 c->fdct = ff_jpeg_fdct_islow; //slow/accurate/default
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
4551 c->fdct248 = ff_fdct248_islow;
1567
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1557
diff changeset
4552 }
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4553 #endif //CONFIG_ENCODERS
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4554
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4555 if(avctx->lowres==1){
8596
68e959302527 replace all occurrence of ENABLE_ by the corresponding CONFIG_, HAVE_ or ARCH_
aurel
parents: 8590
diff changeset
4556 if(avctx->idct_algo==FF_IDCT_INT || avctx->idct_algo==FF_IDCT_AUTO || !CONFIG_H264_DECODER){
2272
cd43603c46f9 move h264 idct to its own file and call via function pointer in DspContext
michael
parents: 2259
diff changeset
4557 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
4558 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
4559 }else{
cd43603c46f9 move h264 idct to its own file and call via function pointer in DspContext
michael
parents: 2259
diff changeset
4560 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
4561 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
4562 }
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4563 c->idct = j_rev_dct4;
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4564 c->idct_permutation_type= FF_NO_IDCT_PERM;
2257
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
4565 }else if(avctx->lowres==2){
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
4566 c->idct_put= ff_jref_idct2_put;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
4567 c->idct_add= ff_jref_idct2_add;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
4568 c->idct = j_rev_dct2;
5f64a30339e5 1/4 resolution decoding
michael
parents: 2256
diff changeset
4569 c->idct_permutation_type= FF_NO_IDCT_PERM;
2259
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
4570 }else if(avctx->lowres==3){
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
4571 c->idct_put= ff_jref_idct1_put;
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
4572 c->idct_add= ff_jref_idct1_add;
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
4573 c->idct = j_rev_dct1;
12e75af1d44c 1/8 resolution decoding
michael
parents: 2257
diff changeset
4574 c->idct_permutation_type= FF_NO_IDCT_PERM;
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4575 }else{
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4576 if(avctx->idct_algo==FF_IDCT_INT){
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4577 c->idct_put= ff_jref_idct_put;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4578 c->idct_add= ff_jref_idct_add;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4579 c->idct = j_rev_dct;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4580 c->idct_permutation_type= FF_LIBMPEG2_IDCT_PERM;
9975
d6d7e8d4a04d Do not redundantly check for both CONFIG_THEORA_DECODER and CONFIG_VP3_DECODER.
diego
parents: 9586
diff changeset
4581 }else if((CONFIG_VP3_DECODER || CONFIG_VP5_DECODER || CONFIG_VP6_DECODER ) &&
5007
f7edc4fe94db Make vp3dsp*.c compilation optional.
takis
parents: 4988
diff changeset
4582 avctx->idct_algo==FF_IDCT_VP3){
2693
02925a3903b6 porting vp3 idct over to lavc idct api
michael
parents: 2651
diff changeset
4583 c->idct_put= ff_vp3_idct_put_c;
02925a3903b6 porting vp3 idct over to lavc idct api
michael
parents: 2651
diff changeset
4584 c->idct_add= ff_vp3_idct_add_c;
02925a3903b6 porting vp3 idct over to lavc idct api
michael
parents: 2651
diff changeset
4585 c->idct = ff_vp3_idct_c;
02925a3903b6 porting vp3 idct over to lavc idct api
michael
parents: 2651
diff changeset
4586 c->idct_permutation_type= FF_NO_IDCT_PERM;
5887
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4587 }else if(avctx->idct_algo==FF_IDCT_WMV2){
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4588 c->idct_put= ff_wmv2_idct_put_c;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4589 c->idct_add= ff_wmv2_idct_add_c;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4590 c->idct = ff_wmv2_idct_c;
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4591 c->idct_permutation_type= FF_NO_IDCT_PERM;
6407
7f9e2b5893fc floating point AAN IDCT
michael
parents: 6387
diff changeset
4592 }else if(avctx->idct_algo==FF_IDCT_FAAN){
7f9e2b5893fc floating point AAN IDCT
michael
parents: 6387
diff changeset
4593 c->idct_put= ff_faanidct_put;
7f9e2b5893fc floating point AAN IDCT
michael
parents: 6387
diff changeset
4594 c->idct_add= ff_faanidct_add;
7f9e2b5893fc floating point AAN IDCT
michael
parents: 6387
diff changeset
4595 c->idct = ff_faanidct;
7f9e2b5893fc floating point AAN IDCT
michael
parents: 6387
diff changeset
4596 c->idct_permutation_type= FF_NO_IDCT_PERM;
8596
68e959302527 replace all occurrence of ENABLE_ by the corresponding CONFIG_, HAVE_ or ARCH_
aurel
parents: 8590
diff changeset
4597 }else if(CONFIG_EATGQ_DECODER && avctx->idct_algo==FF_IDCT_EA) {
8120
3d4e01bcd2a5 Electronic Arts TGQ/TQI/MAD IDCT algorithm
pross
parents: 7995
diff changeset
4598 c->idct_put= ff_ea_idct_put_c;
3d4e01bcd2a5 Electronic Arts TGQ/TQI/MAD IDCT algorithm
pross
parents: 7995
diff changeset
4599 c->idct_permutation_type= FF_NO_IDCT_PERM;
11231
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
4600 }else if(CONFIG_BINK_DECODER && avctx->idct_algo==FF_IDCT_BINK) {
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
4601 c->idct = ff_bink_idct_c;
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
4602 c->idct_add = ff_bink_idct_add_c;
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
4603 c->idct_put = ff_bink_idct_put_c;
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
4604 c->idct_permutation_type = FF_NO_IDCT_PERM;
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4605 }else{ //accurate/default
6001
f4859c13426b add ff_ prefix to all simple_idct symbols
aurel
parents: 5899
diff changeset
4606 c->idct_put= ff_simple_idct_put;
f4859c13426b add ff_ prefix to all simple_idct symbols
aurel
parents: 5899
diff changeset
4607 c->idct_add= ff_simple_idct_add;
f4859c13426b add ff_ prefix to all simple_idct symbols
aurel
parents: 5899
diff changeset
4608 c->idct = ff_simple_idct;
2256
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4609 c->idct_permutation_type= FF_NO_IDCT_PERM;
7e0b2e86afa9 1/2 resolution decoding
michael
parents: 2208
diff changeset
4610 }
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4611 }
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4612
8596
68e959302527 replace all occurrence of ENABLE_ by the corresponding CONFIG_, HAVE_ or ARCH_
aurel
parents: 8590
diff changeset
4613 if (CONFIG_H264_DECODER) {
5065
d0f834073f6b Reindent
takis
parents: 5064
diff changeset
4614 c->h264_idct_add= ff_h264_idct_add_c;
d0f834073f6b Reindent
takis
parents: 5064
diff changeset
4615 c->h264_idct8_add= ff_h264_idct8_add_c;
d0f834073f6b Reindent
takis
parents: 5064
diff changeset
4616 c->h264_idct_dc_add= ff_h264_idct_dc_add_c;
d0f834073f6b Reindent
takis
parents: 5064
diff changeset
4617 c->h264_idct8_dc_add= ff_h264_idct8_dc_add_c;
8375
de2509cf3c44 H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents: 8359
diff changeset
4618 c->h264_idct_add16 = ff_h264_idct_add16_c;
de2509cf3c44 H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents: 8359
diff changeset
4619 c->h264_idct8_add4 = ff_h264_idct8_add4_c;
de2509cf3c44 H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents: 8359
diff changeset
4620 c->h264_idct_add8 = ff_h264_idct_add8_c;
de2509cf3c44 H.264 idct functions that include the chroma, inter luma and intra16 luma loops
michael
parents: 8359
diff changeset
4621 c->h264_idct_add16intra= ff_h264_idct_add16intra_c;
5064
670618daaeb1 Make the compilation of h264idct.c optional.
takis
parents: 5007
diff changeset
4622 }
2272
cd43603c46f9 move h264 idct to its own file and call via function pointer in DspContext
michael
parents: 2259
diff changeset
4623
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
4624 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
4625 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
4626 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
4627 c->put_signed_pixels_clamped = put_signed_pixels_clamped_c;
11231
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
4628 c->put_pixels_nonclamped = put_pixels_nonclamped_c;
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
4629 c->add_pixels_clamped = add_pixels_clamped_c;
2763
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
4630 c->add_pixels8 = add_pixels8_c;
01ed2c3b8f08 H.264 lossless mode
lorenm
parents: 2755
diff changeset
4631 c->add_pixels4 = add_pixels4_c;
4988
689490842cf5 factor sum_abs_dctelem out of dct_sad, and simd it.
lorenm
parents: 4749
diff changeset
4632 c->sum_abs_dctelem = sum_abs_dctelem_c;
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
4633 c->gmc1 = gmc1_c;
3248
7aa9f80e7954 mmx implementation of 3-point GMC. (5x faster than C)
lorenm
parents: 3245
diff changeset
4634 c->gmc = ff_gmc_c;
8288
800444234375 clear_block mmx
lorenm
parents: 8232
diff changeset
4635 c->clear_block = clear_block_c;
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 764
diff changeset
4636 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
4637 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
4638 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
4639
11231
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
4640 c->fill_block_tab[0] = fill_block16_c;
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
4641 c->fill_block_tab[1] = fill_block8_c;
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
4642 c->scale_block = scale_block_c;
0fc1cdd984b7 Bink video decoder
kostya
parents: 11195
diff changeset
4643
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4644 /* TODO [0] 16 [1] 8 */
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4645 c->pix_abs[0][0] = pix_abs16_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4646 c->pix_abs[0][1] = pix_abs16_x2_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4647 c->pix_abs[0][2] = pix_abs16_y2_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4648 c->pix_abs[0][3] = pix_abs16_xy2_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4649 c->pix_abs[1][0] = pix_abs8_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4650 c->pix_abs[1][1] = pix_abs8_x2_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4651 c->pix_abs[1][2] = pix_abs8_y2_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4652 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
4653
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4654 #define dspfunc(PFX, IDX, NUM) \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4655 c->PFX ## _pixels_tab[IDX][0] = PFX ## _pixels ## NUM ## _c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4656 c->PFX ## _pixels_tab[IDX][1] = PFX ## _pixels ## NUM ## _x2_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4657 c->PFX ## _pixels_tab[IDX][2] = PFX ## _pixels ## NUM ## _y2_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4658 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
4659
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4660 dspfunc(put, 0, 16);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4661 dspfunc(put_no_rnd, 0, 16);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4662 dspfunc(put, 1, 8);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4663 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
4664 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
4665 dspfunc(put, 3, 2);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
4666
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4667 dspfunc(avg, 0, 16);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4668 dspfunc(avg_no_rnd, 0, 16);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4669 dspfunc(avg, 1, 8);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4670 dspfunc(avg_no_rnd, 1, 8);
1319
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
4671 dspfunc(avg, 2, 4);
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
4672 dspfunc(avg, 3, 2);
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4673 #undef dspfunc
857
b510a7b6decd 100l for kabi (fixing qpel)
michaelni
parents: 853
diff changeset
4674
1864
9a73e6dc5d68 fix motion compensation with (x+1/2,y+1/2) MVs
michael
parents: 1784
diff changeset
4675 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
4676 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
4677
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
4678 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
4679 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
4680 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
4681 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
4682 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
4683 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
4684 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
4685 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
4686 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
4687
1319
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
4688 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
4689 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
4690 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
4691 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
4692 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
4693 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
4694 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
4695 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
4696 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
4697
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4698 #define dspfunc(PFX, IDX, NUM) \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4699 c->PFX ## _pixels_tab[IDX][ 0] = PFX ## NUM ## _mc00_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4700 c->PFX ## _pixels_tab[IDX][ 1] = PFX ## NUM ## _mc10_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4701 c->PFX ## _pixels_tab[IDX][ 2] = PFX ## NUM ## _mc20_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4702 c->PFX ## _pixels_tab[IDX][ 3] = PFX ## NUM ## _mc30_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4703 c->PFX ## _pixels_tab[IDX][ 4] = PFX ## NUM ## _mc01_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4704 c->PFX ## _pixels_tab[IDX][ 5] = PFX ## NUM ## _mc11_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4705 c->PFX ## _pixels_tab[IDX][ 6] = PFX ## NUM ## _mc21_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4706 c->PFX ## _pixels_tab[IDX][ 7] = PFX ## NUM ## _mc31_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4707 c->PFX ## _pixels_tab[IDX][ 8] = PFX ## NUM ## _mc02_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4708 c->PFX ## _pixels_tab[IDX][ 9] = PFX ## NUM ## _mc12_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4709 c->PFX ## _pixels_tab[IDX][10] = PFX ## NUM ## _mc22_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4710 c->PFX ## _pixels_tab[IDX][11] = PFX ## NUM ## _mc32_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4711 c->PFX ## _pixels_tab[IDX][12] = PFX ## NUM ## _mc03_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4712 c->PFX ## _pixels_tab[IDX][13] = PFX ## NUM ## _mc13_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4713 c->PFX ## _pixels_tab[IDX][14] = PFX ## NUM ## _mc23_c; \
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4714 c->PFX ## _pixels_tab[IDX][15] = PFX ## NUM ## _mc33_c
857
b510a7b6decd 100l for kabi (fixing qpel)
michaelni
parents: 853
diff changeset
4715
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4716 dspfunc(put_qpel, 0, 16);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4717 dspfunc(put_no_rnd_qpel, 0, 16);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4718
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4719 dspfunc(avg_qpel, 0, 16);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4720 /* dspfunc(avg_no_rnd_qpel, 0, 16); */
857
b510a7b6decd 100l for kabi (fixing qpel)
michaelni
parents: 853
diff changeset
4721
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4722 dspfunc(put_qpel, 1, 8);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4723 dspfunc(put_no_rnd_qpel, 1, 8);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4724
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4725 dspfunc(avg_qpel, 1, 8);
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4726 /* dspfunc(avg_no_rnd_qpel, 1, 8); */
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4727
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4728 dspfunc(put_h264_qpel, 0, 16);
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4729 dspfunc(put_h264_qpel, 1, 8);
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4730 dspfunc(put_h264_qpel, 2, 4);
3020
c75fb0747e74 use h264 MC functions for 2xX Xx2 blocks in snow too
michael
parents: 3013
diff changeset
4731 dspfunc(put_h264_qpel, 3, 2);
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4732 dspfunc(avg_h264_qpel, 0, 16);
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4733 dspfunc(avg_h264_qpel, 1, 8);
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4734 dspfunc(avg_h264_qpel, 2, 4);
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4735
859
9512471dec40 * using macros to shorten init part
kabi
parents: 858
diff changeset
4736 #undef dspfunc
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4737 c->put_h264_chroma_pixels_tab[0]= put_h264_chroma_mc8_c;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4738 c->put_h264_chroma_pixels_tab[1]= put_h264_chroma_mc4_c;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4739 c->put_h264_chroma_pixels_tab[2]= put_h264_chroma_mc2_c;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4740 c->avg_h264_chroma_pixels_tab[0]= avg_h264_chroma_mc8_c;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4741 c->avg_h264_chroma_pixels_tab[1]= avg_h264_chroma_mc4_c;
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1106
diff changeset
4742 c->avg_h264_chroma_pixels_tab[2]= avg_h264_chroma_mc2_c;
9439
ef3a7b711cc0 Rename put_no_rnd_h264_chroma* to reflect its usage in VC1 only
conrad
parents: 9437
diff changeset
4743 c->put_no_rnd_vc1_chroma_pixels_tab[0]= put_no_rnd_vc1_chroma_mc8_c;
9440
daee921fb6bb VC1: add and use avg_no_rnd chroma MC functions
conrad
parents: 9439
diff changeset
4744 c->avg_no_rnd_vc1_chroma_pixels_tab[0]= avg_no_rnd_vc1_chroma_mc8_c;
857
b510a7b6decd 100l for kabi (fixing qpel)
michaelni
parents: 853
diff changeset
4745
2415
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4746 c->weight_h264_pixels_tab[0]= weight_h264_pixels16x16_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4747 c->weight_h264_pixels_tab[1]= weight_h264_pixels16x8_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4748 c->weight_h264_pixels_tab[2]= weight_h264_pixels8x16_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4749 c->weight_h264_pixels_tab[3]= weight_h264_pixels8x8_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4750 c->weight_h264_pixels_tab[4]= weight_h264_pixels8x4_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4751 c->weight_h264_pixels_tab[5]= weight_h264_pixels4x8_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4752 c->weight_h264_pixels_tab[6]= weight_h264_pixels4x4_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4753 c->weight_h264_pixels_tab[7]= weight_h264_pixels4x2_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4754 c->weight_h264_pixels_tab[8]= weight_h264_pixels2x4_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4755 c->weight_h264_pixels_tab[9]= weight_h264_pixels2x2_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4756 c->biweight_h264_pixels_tab[0]= biweight_h264_pixels16x16_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4757 c->biweight_h264_pixels_tab[1]= biweight_h264_pixels16x8_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4758 c->biweight_h264_pixels_tab[2]= biweight_h264_pixels8x16_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4759 c->biweight_h264_pixels_tab[3]= biweight_h264_pixels8x8_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4760 c->biweight_h264_pixels_tab[4]= biweight_h264_pixels8x4_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4761 c->biweight_h264_pixels_tab[5]= biweight_h264_pixels4x8_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4762 c->biweight_h264_pixels_tab[6]= biweight_h264_pixels4x4_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4763 c->biweight_h264_pixels_tab[7]= biweight_h264_pixels4x2_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4764 c->biweight_h264_pixels_tab[8]= biweight_h264_pixels2x4_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4765 c->biweight_h264_pixels_tab[9]= biweight_h264_pixels2x2_c;
db2cf6005d19 H.264 weighted prediction.
lorenm
parents: 2382
diff changeset
4766
6437
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
4767 c->draw_edges = draw_edges_c;
5154ab444372 move draw_edges() into dsputil
aurel
parents: 6407
diff changeset
4768
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8496
diff changeset
4769 #if CONFIG_CAVS_DECODER
3395
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
4770 ff_cavsdsp_init(c,avctx);
3432
ab49baf4adad Fix compilation with --disable-decoders.
diego
parents: 3395
diff changeset
4771 #endif
9585
5e1d9508b62f mlpdec: Move MLP's filter_channel() to dsputils.
ramiro
parents: 9440
diff changeset
4772
5e1d9508b62f mlpdec: Move MLP's filter_channel() to dsputils.
ramiro
parents: 9440
diff changeset
4773 #if CONFIG_MLP_DECODER || CONFIG_TRUEHD_DECODER
5e1d9508b62f mlpdec: Move MLP's filter_channel() to dsputils.
ramiro
parents: 9440
diff changeset
4774 ff_mlp_init(c, avctx);
5e1d9508b62f mlpdec: Move MLP's filter_channel() to dsputils.
ramiro
parents: 9440
diff changeset
4775 #endif
9995
3141f69e3905 Do not check for both CONFIG_VC1_DECODER and CONFIG_WMV3_DECODER,
diego
parents: 9975
diff changeset
4776 #if CONFIG_VC1_DECODER
3526
7dc8e4a12105 New functions in DSPContext for VC-1 decoding
kostya
parents: 3432
diff changeset
4777 ff_vc1dsp_init(c,avctx);
7dc8e4a12105 New functions in DSPContext for VC-1 decoding
kostya
parents: 3432
diff changeset
4778 #endif
9995
3141f69e3905 Do not check for both CONFIG_VC1_DECODER and CONFIG_WMV3_DECODER,
diego
parents: 9975
diff changeset
4779 #if CONFIG_WMV2_DECODER || CONFIG_VC1_DECODER
5887
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4780 ff_intrax8dsp_init(c,avctx);
83ac4620c6ed intrax8 decoder patch by "someone"
michael
parents: 5737
diff changeset
4781 #endif
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8496
diff changeset
4782 #if CONFIG_RV30_DECODER
8410
b0ec9768b39b RV30 decoder
kostya
parents: 8395
diff changeset
4783 ff_rv30dsp_init(c,avctx);
b0ec9768b39b RV30 decoder
kostya
parents: 8395
diff changeset
4784 #endif
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8496
diff changeset
4785 #if CONFIG_RV40_DECODER
8232
520a6e56b8c0 RV40 MC functions
kostya
parents: 8120
diff changeset
4786 ff_rv40dsp_init(c,avctx);
520a6e56b8c0 RV40 MC functions
kostya
parents: 8120
diff changeset
4787 c->put_rv40_qpel_pixels_tab[0][15] = put_rv40_qpel16_mc33_c;
520a6e56b8c0 RV40 MC functions
kostya
parents: 8120
diff changeset
4788 c->avg_rv40_qpel_pixels_tab[0][15] = avg_rv40_qpel16_mc33_c;
520a6e56b8c0 RV40 MC functions
kostya
parents: 8120
diff changeset
4789 c->put_rv40_qpel_pixels_tab[1][15] = put_rv40_qpel8_mc33_c;
520a6e56b8c0 RV40 MC functions
kostya
parents: 8120
diff changeset
4790 c->avg_rv40_qpel_pixels_tab[1][15] = avg_rv40_qpel8_mc33_c;
520a6e56b8c0 RV40 MC functions
kostya
parents: 8120
diff changeset
4791 #endif
3395
adccbf4a1040 CAVS decoder by (Stefan Gehrer stefan.gehrer gmx.de)
michael
parents: 3373
diff changeset
4792
936
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
4793 c->put_mspel_pixels_tab[0]= put_mspel8_mc00_c;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
4794 c->put_mspel_pixels_tab[1]= put_mspel8_mc10_c;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
4795 c->put_mspel_pixels_tab[2]= put_mspel8_mc20_c;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
4796 c->put_mspel_pixels_tab[3]= put_mspel8_mc30_c;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
4797 c->put_mspel_pixels_tab[4]= put_mspel8_mc02_c;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
4798 c->put_mspel_pixels_tab[5]= put_mspel8_mc12_c;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
4799 c->put_mspel_pixels_tab[6]= put_mspel8_mc22_c;
caa77cd960c0 qpel encoding
michaelni
parents: 909
diff changeset
4800 c->put_mspel_pixels_tab[7]= put_mspel8_mc32_c;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4801
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4802 #define SET_CMP_FUNC(name) \
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4803 c->name[0]= name ## 16_c;\
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4804 c->name[1]= name ## 8x8_c;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4805
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4806 SET_CMP_FUNC(hadamard8_diff)
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4807 c->hadamard8_diff[4]= hadamard8_intra16_c;
8978
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4808 c->hadamard8_diff[5]= hadamard8_intra8x8_c;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4809 SET_CMP_FUNC(dct_sad)
2382
23e7af611c34 more flexible frame skip decission
michael
parents: 2272
diff changeset
4810 SET_CMP_FUNC(dct_max)
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8496
diff changeset
4811 #if CONFIG_GPL
3010
533c6386eca9 8x8 integer dct from x264 as cmp function (under CONFIG_GPL)
michael
parents: 2979
diff changeset
4812 SET_CMP_FUNC(dct264_sad)
3013
b6a90fabaa76 ugly, but necessary
melanson
parents: 3010
diff changeset
4813 #endif
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4814 c->sad[0]= pix_abs16_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4815 c->sad[1]= pix_abs8_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4816 c->sse[0]= sse16_c;
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4817 c->sse[1]= sse8_c;
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
4818 c->sse[2]= sse4_c;
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4819 SET_CMP_FUNC(quant_psnr)
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4820 SET_CMP_FUNC(rd)
dea5b2946999 interlaced motion estimation
michael
parents: 1689
diff changeset
4821 SET_CMP_FUNC(bit)
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4822 c->vsad[0]= vsad16_c;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4823 c->vsad[4]= vsad_intra16_c;
8978
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4824 c->vsad[5]= vsad_intra8_c;
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4825 c->vsse[0]= vsse16_c;
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1708
diff changeset
4826 c->vsse[4]= vsse_intra16_c;
8978
a49197cd37ce Adding 2 intra 8x8 cmp functions: vsad, vsse
romansh
parents: 8976
diff changeset
4827 c->vsse[5]= vsse_intra8_c;
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
4828 c->nsse[0]= nsse16_c;
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
4829 c->nsse[1]= nsse8_c;
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8496
diff changeset
4830 #if CONFIG_SNOW_ENCODER
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
4831 c->w53[0]= w53_16_c;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
4832 c->w53[1]= w53_8_c;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
4833 c->w97[0]= w97_16_c;
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
4834 c->w97[1]= w97_8_c;
3373
b8996cc5ccae Disable w53 and w97 cmp methods when snow encoder is disabled
gpoirier
parents: 3323
diff changeset
4835 #endif
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2169
diff changeset
4836
4749
7011f597e473 mmx 16-bit ssd. 2.3x faster svq1 encoding.
lorenm
parents: 4598
diff changeset
4837 c->ssd_int8_vs_int16 = ssd_int8_vs_int16_c;
7011f597e473 mmx 16-bit ssd. 2.3x faster svq1 encoding.
lorenm
parents: 4598
diff changeset
4838
866
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
4839 c->add_bytes= add_bytes_c;
6384
0a403ade8c81 simd and unroll png_filter_row
lorenm
parents: 6241
diff changeset
4840 c->add_bytes_l2= add_bytes_l2_c;
866
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
4841 c->diff_bytes= diff_bytes_c;
8760
31138c296ac6 ff_add_hfyu_median_prediction_mmx2
lorenm
parents: 8719
diff changeset
4842 c->add_hfyu_median_prediction= add_hfyu_median_prediction_c;
1527
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1329
diff changeset
4843 c->sub_hfyu_median_prediction= sub_hfyu_median_prediction_c;
10370
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
4844 c->add_hfyu_left_prediction = add_hfyu_left_prediction_c;
6009feb54020 Move HuffYUV left prediction to dsputil.
cehoyos
parents: 10300
diff changeset
4845 c->add_hfyu_left_prediction_bgr32 = add_hfyu_left_prediction_bgr32_c;
1273
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
4846 c->bswap_buf= bswap_buf;
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8496
diff changeset
4847 #if CONFIG_PNG_DECODER
6384
0a403ade8c81 simd and unroll png_filter_row
lorenm
parents: 6241
diff changeset
4848 c->add_png_paeth_prediction= ff_add_png_paeth_prediction;
0a403ade8c81 simd and unroll png_filter_row
lorenm
parents: 6241
diff changeset
4849 #endif
2633
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
4850
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
4851 c->h264_v_loop_filter_luma= h264_v_loop_filter_luma_c;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
4852 c->h264_h_loop_filter_luma= h264_h_loop_filter_luma_c;
8395
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
4853 c->h264_v_loop_filter_luma_intra= h264_v_loop_filter_luma_intra_c;
195cba8f6257 Move filter_luma_intra into dsputil for later addition of asm.
darkshikari
parents: 8375
diff changeset
4854 c->h264_h_loop_filter_luma_intra= h264_h_loop_filter_luma_intra_c;
2633
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
4855 c->h264_v_loop_filter_chroma= h264_v_loop_filter_chroma_c;
72e6ffa1f3a5 MMX for H.264 deblocking filter
lorenm
parents: 2522
diff changeset
4856 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
4857 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
4858 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
4859 c->h264_loop_filter_strength= NULL;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4860
10749
5cca4b6c459d Get rid of pointless CONFIG_ANY_H263 preprocessor definition.
diego
parents: 10748
diff changeset
4861 if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
5278
ef85411bb7e8 cosmetics: indentation
aurel
parents: 5277
diff changeset
4862 c->h263_h_loop_filter= h263_h_loop_filter_c;
ef85411bb7e8 cosmetics: indentation
aurel
parents: 5277
diff changeset
4863 c->h263_v_loop_filter= h263_v_loop_filter_c;
5277
7b3fcb7c61ce Avoid linking with h263.c functions when the relevant codecs
aurel
parents: 5256
diff changeset
4864 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4865
9975
d6d7e8d4a04d Do not redundantly check for both CONFIG_THEORA_DECODER and CONFIG_VP3_DECODER.
diego
parents: 9586
diff changeset
4866 if (CONFIG_VP3_DECODER) {
7995
1fbfce20cb79 Move VP3 loop filter to DSPContext
conrad
parents: 7564
diff changeset
4867 c->vp3_h_loop_filter= ff_vp3_h_loop_filter_c;
1fbfce20cb79 Move VP3 loop filter to DSPContext
conrad
parents: 7564
diff changeset
4868 c->vp3_v_loop_filter= ff_vp3_v_loop_filter_c;
1fbfce20cb79 Move VP3 loop filter to DSPContext
conrad
parents: 7564
diff changeset
4869 }
8785
bee83b3f9a6b move vp6_filter_diag4() to a new vp6dsp.c file and use it throught dsputil
aurel
parents: 8760
diff changeset
4870 if (CONFIG_VP6_DECODER) {
bee83b3f9a6b move vp6_filter_diag4() to a new vp6dsp.c file and use it throught dsputil
aurel
parents: 8760
diff changeset
4871 c->vp6_filter_diag4= ff_vp6_filter_diag4_c;
bee83b3f9a6b move vp6_filter_diag4() to a new vp6dsp.c file and use it throught dsputil
aurel
parents: 8760
diff changeset
4872 }
7995
1fbfce20cb79 Move VP3 loop filter to DSPContext
conrad
parents: 7564
diff changeset
4873
2045
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
4874 c->h261_loop_filter= h261_loop_filter_c;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
4875
1784
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
4876 c->try_8x8basis= try_8x8basis_c;
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1739
diff changeset
4877 c->add_8x8basis= add_8x8basis_c;
866
725ef4ea3ecc huffyuv
michaelni
parents: 861
diff changeset
4878
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8496
diff changeset
4879 #if CONFIG_SNOW_DECODER
3198
6b9f0c4fbdbe First part of a series of speed-enchancing patches.
gpoirier
parents: 3105
diff changeset
4880 c->vertical_compose97i = ff_snow_vertical_compose97i;
6b9f0c4fbdbe First part of a series of speed-enchancing patches.
gpoirier
parents: 3105
diff changeset
4881 c->horizontal_compose97i = ff_snow_horizontal_compose97i;
6b9f0c4fbdbe First part of a series of speed-enchancing patches.
gpoirier
parents: 3105
diff changeset
4882 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
4883 #endif
3198
6b9f0c4fbdbe First part of a series of speed-enchancing patches.
gpoirier
parents: 3105
diff changeset
4884
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8496
diff changeset
4885 #if CONFIG_VORBIS_DECODER
3536
545a15c19c91 sse & sse2 implementations of vorbis channel coupling.
lorenm
parents: 3526
diff changeset
4886 c->vorbis_inverse_coupling = vorbis_inverse_coupling;
545a15c19c91 sse & sse2 implementations of vorbis channel coupling.
lorenm
parents: 3526
diff changeset
4887 #endif
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 8496
diff changeset
4888 #if CONFIG_AC3_DECODER
7563
8390efaa0c03 simd downmix
lorenm
parents: 7286
diff changeset
4889 c->ac3_downmix = ff_ac3_downmix_c;
8390efaa0c03 simd downmix
lorenm
parents: 7286
diff changeset
4890 #endif
10429
289dd8daf4ee add CONFIG_LPC to the build system for lpc dsputil functions. fixes build
jbr
parents: 10424
diff changeset
4891 #if CONFIG_LPC
10424
94595d0e617c Move autocorrelation function from flacenc.c to lpc.c. Also rename the
jbr
parents: 10421
diff changeset
4892 c->lpc_compute_autocorr = ff_lpc_compute_autocorr;
10429
289dd8daf4ee add CONFIG_LPC to the build system for lpc dsputil functions. fixes build
jbr
parents: 10424
diff changeset
4893 #endif
3568
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4894 c->vector_fmul = vector_fmul_c;
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4895 c->vector_fmul_reverse = vector_fmul_reverse_c;
10300
4d1b9ca628fc Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents: 10219
diff changeset
4896 c->vector_fmul_add = vector_fmul_add_c;
7261
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4897 c->vector_fmul_window = ff_vector_fmul_window_c;
7564
7cf793954871 simd int->float
lorenm
parents: 7563
diff changeset
4898 c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_c;
10104
0fa3d21b317e SSE optimized vector_clipf(). 10% faster TwinVQ decoding.
vitor
parents: 10094
diff changeset
4899 c->vector_clipf = vector_clipf_c;
3568
945caa35ee9a sse and 3dnow implementations of float->int conversion and mdct windowing.
lorenm
parents: 3536
diff changeset
4900 c->float_to_int16 = ff_float_to_int16_c;
7261
032a49f033e8 simplify vorbis windowing
lorenm
parents: 7218
diff changeset
4901 c->float_to_int16_interleave = ff_float_to_int16_interleave_c;
7203
87b1dfb5a98d Add several vector functions used by Monkey's Audio decoder to dsputil
kostya
parents: 6719
diff changeset
4902 c->scalarproduct_int16 = scalarproduct_int16_c;
10644
5da7180afadf refactor and optimize scalarproduct
lorenm
parents: 10431
diff changeset
4903 c->scalarproduct_and_madd_int16 = scalarproduct_and_madd_int16_c;
10219
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4904 c->scalarproduct_float = scalarproduct_float_c;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4905 c->butterflies_float = butterflies_float_c;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4906 c->vector_fmul_scalar = vector_fmul_scalar_c;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4907
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4908 c->vector_fmul_sv_scalar[0] = vector_fmul_sv_scalar_2_c;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4909 c->vector_fmul_sv_scalar[1] = vector_fmul_sv_scalar_4_c;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4910
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4911 c->sv_fmul_scalar[0] = sv_fmul_scalar_2_c;
87ae4a167138 Add some dsputil functions useful for AAC decoder
mru
parents: 10207
diff changeset
4912 c->sv_fmul_scalar[1] = sv_fmul_scalar_4_c;
3536
545a15c19c91 sse & sse2 implementations of vorbis channel coupling.
lorenm
parents: 3526
diff changeset
4913
3245
c2c29be6282e remove mpegvideo.c img resample dependancy
michael
parents: 3215
diff changeset
4914 c->shrink[0]= ff_img_copy_plane;
c2c29be6282e remove mpegvideo.c img resample dependancy
michael
parents: 3215
diff changeset
4915 c->shrink[1]= ff_shrink22;
c2c29be6282e remove mpegvideo.c img resample dependancy
michael
parents: 3215
diff changeset
4916 c->shrink[2]= ff_shrink44;
c2c29be6282e remove mpegvideo.c img resample dependancy
michael
parents: 3215
diff changeset
4917 c->shrink[3]= ff_shrink88;
c2c29be6282e remove mpegvideo.c img resample dependancy
michael
parents: 3215
diff changeset
4918
3215
06f98047ff26 prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents: 3199
diff changeset
4919 c->prefetch= just_return;
06f98047ff26 prefetch pixels for future motion compensation. 2-5% faster h264.
lorenm
parents: 3199
diff changeset
4920
3807
6a40092eb9e6 approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents: 3728
diff changeset
4921 memset(c->put_2tap_qpel_pixels_tab, 0, sizeof(c->put_2tap_qpel_pixels_tab));
6a40092eb9e6 approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents: 3728
diff changeset
4922 memset(c->avg_2tap_qpel_pixels_tab, 0, sizeof(c->avg_2tap_qpel_pixels_tab));
6a40092eb9e6 approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents: 3728
diff changeset
4923
8596
68e959302527 replace all occurrence of ENABLE_ by the corresponding CONFIG_, HAVE_ or ARCH_
aurel
parents: 8590
diff changeset
4924 if (HAVE_MMX) dsputil_init_mmx (c, avctx);
68e959302527 replace all occurrence of ENABLE_ by the corresponding CONFIG_, HAVE_ or ARCH_
aurel
parents: 8590
diff changeset
4925 if (ARCH_ARM) dsputil_init_arm (c, avctx);
68e959302527 replace all occurrence of ENABLE_ by the corresponding CONFIG_, HAVE_ or ARCH_
aurel
parents: 8590
diff changeset
4926 if (CONFIG_MLIB) dsputil_init_mlib (c, avctx);
68e959302527 replace all occurrence of ENABLE_ by the corresponding CONFIG_, HAVE_ or ARCH_
aurel
parents: 8590
diff changeset
4927 if (HAVE_VIS) dsputil_init_vis (c, avctx);
68e959302527 replace all occurrence of ENABLE_ by the corresponding CONFIG_, HAVE_ or ARCH_
aurel
parents: 8590
diff changeset
4928 if (ARCH_ALPHA) dsputil_init_alpha (c, avctx);
68e959302527 replace all occurrence of ENABLE_ by the corresponding CONFIG_, HAVE_ or ARCH_
aurel
parents: 8590
diff changeset
4929 if (ARCH_PPC) dsputil_init_ppc (c, avctx);
68e959302527 replace all occurrence of ENABLE_ by the corresponding CONFIG_, HAVE_ or ARCH_
aurel
parents: 8590
diff changeset
4930 if (HAVE_MMI) dsputil_init_mmi (c, avctx);
68e959302527 replace all occurrence of ENABLE_ by the corresponding CONFIG_, HAVE_ or ARCH_
aurel
parents: 8590
diff changeset
4931 if (ARCH_SH4) dsputil_init_sh4 (c, avctx);
68e959302527 replace all occurrence of ENABLE_ by the corresponding CONFIG_, HAVE_ or ARCH_
aurel
parents: 8590
diff changeset
4932 if (ARCH_BFIN) dsputil_init_bfin (c, avctx);
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4933
3807
6a40092eb9e6 approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents: 3728
diff changeset
4934 for(i=0; i<64; i++){
6a40092eb9e6 approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents: 3728
diff changeset
4935 if(!c->put_2tap_qpel_pixels_tab[0][i])
6a40092eb9e6 approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents: 3728
diff changeset
4936 c->put_2tap_qpel_pixels_tab[0][i]= c->put_h264_qpel_pixels_tab[0][i];
6a40092eb9e6 approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents: 3728
diff changeset
4937 if(!c->avg_2tap_qpel_pixels_tab[0][i])
6a40092eb9e6 approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents: 3728
diff changeset
4938 c->avg_2tap_qpel_pixels_tab[0][i]= c->avg_h264_qpel_pixels_tab[0][i];
6a40092eb9e6 approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents: 3728
diff changeset
4939 }
6a40092eb9e6 approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents: 3728
diff changeset
4940
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4941 switch(c->idct_permutation_type){
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4942 case FF_NO_IDCT_PERM:
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4943 for(i=0; i<64; i++)
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4944 c->idct_permutation[i]= i;
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4945 break;
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4946 case FF_LIBMPEG2_IDCT_PERM:
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4947 for(i=0; i<64; i++)
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4948 c->idct_permutation[i]= (i & 0x38) | ((i & 6) >> 1) | ((i & 1) << 2);
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4949 break;
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4950 case FF_SIMPLE_IDCT_PERM:
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4951 for(i=0; i<64; i++)
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4952 c->idct_permutation[i]= simple_mmx_permutation[i];
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4953 break;
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4954 case FF_TRANSPOSE_IDCT_PERM:
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4955 for(i=0; i<64; i++)
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4956 c->idct_permutation[i]= ((i&7)<<3) | (i>>3);
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4957 break;
2696
9699d325049d porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents: 2693
diff changeset
4958 case FF_PARTTRANS_IDCT_PERM:
9699d325049d porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents: 2693
diff changeset
4959 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
4960 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
4961 break;
6600
c3213c91124c Add a new IDCT permutation, used in xvid_sse2 and possibly future similar IDCTs.
astrange
parents: 6450
diff changeset
4962 case FF_SSE2_IDCT_PERM:
c3213c91124c Add a new IDCT permutation, used in xvid_sse2 and possibly future similar IDCTs.
astrange
parents: 6450
diff changeset
4963 for(i=0; i<64; i++)
c3213c91124c Add a new IDCT permutation, used in xvid_sse2 and possibly future similar IDCTs.
astrange
parents: 6450
diff changeset
4964 c->idct_permutation[i]= (i&0x38) | idct_sse2_row_perm[i&7];
c3213c91124c Add a new IDCT permutation, used in xvid_sse2 and possibly future similar IDCTs.
astrange
parents: 6450
diff changeset
4965 break;
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
4966 default:
1598
932d306bf1dc av_log() patch by (Michel Bardiaux <mbardiaux at peaktime dot be>)
michael
parents: 1571
diff changeset
4967 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
4968 }
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
4969 }
252
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 220
diff changeset
4970