yading@10
|
1 /*
|
yading@10
|
2 * ARM NEON optimised DSP functions
|
yading@10
|
3 * Copyright (c) 2008 Mans Rullgard <mans@mansr.com>
|
yading@10
|
4 *
|
yading@10
|
5 * This file is part of FFmpeg.
|
yading@10
|
6 *
|
yading@10
|
7 * FFmpeg is free software; you can redistribute it and/or
|
yading@10
|
8 * modify it under the terms of the GNU Lesser General Public
|
yading@10
|
9 * License as published by the Free Software Foundation; either
|
yading@10
|
10 * version 2.1 of the License, or (at your option) any later version.
|
yading@10
|
11 *
|
yading@10
|
12 * FFmpeg is distributed in the hope that it will be useful,
|
yading@10
|
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
yading@10
|
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
yading@10
|
15 * Lesser General Public License for more details.
|
yading@10
|
16 *
|
yading@10
|
17 * You should have received a copy of the GNU Lesser General Public
|
yading@10
|
18 * License along with FFmpeg; if not, write to the Free Software
|
yading@10
|
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
yading@10
|
20 */
|
yading@10
|
21
|
yading@10
|
22 #include <stdint.h>
|
yading@10
|
23
|
yading@10
|
24 #include "config.h"
|
yading@10
|
25 #include "libavutil/attributes.h"
|
yading@10
|
26 #include "libavutil/arm/cpu.h"
|
yading@10
|
27 #include "libavcodec/h264qpel.h"
|
yading@10
|
28
|
yading@10
|
29 void ff_put_h264_qpel16_mc00_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
30 void ff_put_h264_qpel16_mc10_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
31 void ff_put_h264_qpel16_mc20_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
32 void ff_put_h264_qpel16_mc30_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
33 void ff_put_h264_qpel16_mc01_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
34 void ff_put_h264_qpel16_mc11_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
35 void ff_put_h264_qpel16_mc21_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
36 void ff_put_h264_qpel16_mc31_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
37 void ff_put_h264_qpel16_mc02_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
38 void ff_put_h264_qpel16_mc12_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
39 void ff_put_h264_qpel16_mc22_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
40 void ff_put_h264_qpel16_mc32_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
41 void ff_put_h264_qpel16_mc03_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
42 void ff_put_h264_qpel16_mc13_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
43 void ff_put_h264_qpel16_mc23_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
44 void ff_put_h264_qpel16_mc33_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
45
|
yading@10
|
46 void ff_put_h264_qpel8_mc00_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
47 void ff_put_h264_qpel8_mc10_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
48 void ff_put_h264_qpel8_mc20_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
49 void ff_put_h264_qpel8_mc30_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
50 void ff_put_h264_qpel8_mc01_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
51 void ff_put_h264_qpel8_mc11_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
52 void ff_put_h264_qpel8_mc21_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
53 void ff_put_h264_qpel8_mc31_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
54 void ff_put_h264_qpel8_mc02_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
55 void ff_put_h264_qpel8_mc12_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
56 void ff_put_h264_qpel8_mc22_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
57 void ff_put_h264_qpel8_mc32_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
58 void ff_put_h264_qpel8_mc03_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
59 void ff_put_h264_qpel8_mc13_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
60 void ff_put_h264_qpel8_mc23_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
61 void ff_put_h264_qpel8_mc33_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
62
|
yading@10
|
63 void ff_avg_h264_qpel16_mc00_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
64 void ff_avg_h264_qpel16_mc10_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
65 void ff_avg_h264_qpel16_mc20_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
66 void ff_avg_h264_qpel16_mc30_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
67 void ff_avg_h264_qpel16_mc01_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
68 void ff_avg_h264_qpel16_mc11_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
69 void ff_avg_h264_qpel16_mc21_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
70 void ff_avg_h264_qpel16_mc31_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
71 void ff_avg_h264_qpel16_mc02_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
72 void ff_avg_h264_qpel16_mc12_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
73 void ff_avg_h264_qpel16_mc22_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
74 void ff_avg_h264_qpel16_mc32_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
75 void ff_avg_h264_qpel16_mc03_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
76 void ff_avg_h264_qpel16_mc13_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
77 void ff_avg_h264_qpel16_mc23_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
78 void ff_avg_h264_qpel16_mc33_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
79
|
yading@10
|
80 void ff_avg_h264_qpel8_mc00_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
81 void ff_avg_h264_qpel8_mc10_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
82 void ff_avg_h264_qpel8_mc20_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
83 void ff_avg_h264_qpel8_mc30_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
84 void ff_avg_h264_qpel8_mc01_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
85 void ff_avg_h264_qpel8_mc11_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
86 void ff_avg_h264_qpel8_mc21_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
87 void ff_avg_h264_qpel8_mc31_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
88 void ff_avg_h264_qpel8_mc02_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
89 void ff_avg_h264_qpel8_mc12_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
90 void ff_avg_h264_qpel8_mc22_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
91 void ff_avg_h264_qpel8_mc32_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
92 void ff_avg_h264_qpel8_mc03_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
93 void ff_avg_h264_qpel8_mc13_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
94 void ff_avg_h264_qpel8_mc23_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
95 void ff_avg_h264_qpel8_mc33_neon(uint8_t *, uint8_t *, ptrdiff_t);
|
yading@10
|
96
|
yading@10
|
97 av_cold void ff_h264qpel_init_arm(H264QpelContext *c, int bit_depth)
|
yading@10
|
98 {
|
yading@10
|
99 const int high_bit_depth = bit_depth > 8;
|
yading@10
|
100 int cpu_flags = av_get_cpu_flags();
|
yading@10
|
101
|
yading@10
|
102 if (have_neon(cpu_flags) && !high_bit_depth) {
|
yading@10
|
103 c->put_h264_qpel_pixels_tab[0][ 0] = ff_put_h264_qpel16_mc00_neon;
|
yading@10
|
104 c->put_h264_qpel_pixels_tab[0][ 1] = ff_put_h264_qpel16_mc10_neon;
|
yading@10
|
105 c->put_h264_qpel_pixels_tab[0][ 2] = ff_put_h264_qpel16_mc20_neon;
|
yading@10
|
106 c->put_h264_qpel_pixels_tab[0][ 3] = ff_put_h264_qpel16_mc30_neon;
|
yading@10
|
107 c->put_h264_qpel_pixels_tab[0][ 4] = ff_put_h264_qpel16_mc01_neon;
|
yading@10
|
108 c->put_h264_qpel_pixels_tab[0][ 5] = ff_put_h264_qpel16_mc11_neon;
|
yading@10
|
109 c->put_h264_qpel_pixels_tab[0][ 6] = ff_put_h264_qpel16_mc21_neon;
|
yading@10
|
110 c->put_h264_qpel_pixels_tab[0][ 7] = ff_put_h264_qpel16_mc31_neon;
|
yading@10
|
111 c->put_h264_qpel_pixels_tab[0][ 8] = ff_put_h264_qpel16_mc02_neon;
|
yading@10
|
112 c->put_h264_qpel_pixels_tab[0][ 9] = ff_put_h264_qpel16_mc12_neon;
|
yading@10
|
113 c->put_h264_qpel_pixels_tab[0][10] = ff_put_h264_qpel16_mc22_neon;
|
yading@10
|
114 c->put_h264_qpel_pixels_tab[0][11] = ff_put_h264_qpel16_mc32_neon;
|
yading@10
|
115 c->put_h264_qpel_pixels_tab[0][12] = ff_put_h264_qpel16_mc03_neon;
|
yading@10
|
116 c->put_h264_qpel_pixels_tab[0][13] = ff_put_h264_qpel16_mc13_neon;
|
yading@10
|
117 c->put_h264_qpel_pixels_tab[0][14] = ff_put_h264_qpel16_mc23_neon;
|
yading@10
|
118 c->put_h264_qpel_pixels_tab[0][15] = ff_put_h264_qpel16_mc33_neon;
|
yading@10
|
119
|
yading@10
|
120 c->put_h264_qpel_pixels_tab[1][ 0] = ff_put_h264_qpel8_mc00_neon;
|
yading@10
|
121 c->put_h264_qpel_pixels_tab[1][ 1] = ff_put_h264_qpel8_mc10_neon;
|
yading@10
|
122 c->put_h264_qpel_pixels_tab[1][ 2] = ff_put_h264_qpel8_mc20_neon;
|
yading@10
|
123 c->put_h264_qpel_pixels_tab[1][ 3] = ff_put_h264_qpel8_mc30_neon;
|
yading@10
|
124 c->put_h264_qpel_pixels_tab[1][ 4] = ff_put_h264_qpel8_mc01_neon;
|
yading@10
|
125 c->put_h264_qpel_pixels_tab[1][ 5] = ff_put_h264_qpel8_mc11_neon;
|
yading@10
|
126 c->put_h264_qpel_pixels_tab[1][ 6] = ff_put_h264_qpel8_mc21_neon;
|
yading@10
|
127 c->put_h264_qpel_pixels_tab[1][ 7] = ff_put_h264_qpel8_mc31_neon;
|
yading@10
|
128 c->put_h264_qpel_pixels_tab[1][ 8] = ff_put_h264_qpel8_mc02_neon;
|
yading@10
|
129 c->put_h264_qpel_pixels_tab[1][ 9] = ff_put_h264_qpel8_mc12_neon;
|
yading@10
|
130 c->put_h264_qpel_pixels_tab[1][10] = ff_put_h264_qpel8_mc22_neon;
|
yading@10
|
131 c->put_h264_qpel_pixels_tab[1][11] = ff_put_h264_qpel8_mc32_neon;
|
yading@10
|
132 c->put_h264_qpel_pixels_tab[1][12] = ff_put_h264_qpel8_mc03_neon;
|
yading@10
|
133 c->put_h264_qpel_pixels_tab[1][13] = ff_put_h264_qpel8_mc13_neon;
|
yading@10
|
134 c->put_h264_qpel_pixels_tab[1][14] = ff_put_h264_qpel8_mc23_neon;
|
yading@10
|
135 c->put_h264_qpel_pixels_tab[1][15] = ff_put_h264_qpel8_mc33_neon;
|
yading@10
|
136
|
yading@10
|
137 c->avg_h264_qpel_pixels_tab[0][ 0] = ff_avg_h264_qpel16_mc00_neon;
|
yading@10
|
138 c->avg_h264_qpel_pixels_tab[0][ 1] = ff_avg_h264_qpel16_mc10_neon;
|
yading@10
|
139 c->avg_h264_qpel_pixels_tab[0][ 2] = ff_avg_h264_qpel16_mc20_neon;
|
yading@10
|
140 c->avg_h264_qpel_pixels_tab[0][ 3] = ff_avg_h264_qpel16_mc30_neon;
|
yading@10
|
141 c->avg_h264_qpel_pixels_tab[0][ 4] = ff_avg_h264_qpel16_mc01_neon;
|
yading@10
|
142 c->avg_h264_qpel_pixels_tab[0][ 5] = ff_avg_h264_qpel16_mc11_neon;
|
yading@10
|
143 c->avg_h264_qpel_pixels_tab[0][ 6] = ff_avg_h264_qpel16_mc21_neon;
|
yading@10
|
144 c->avg_h264_qpel_pixels_tab[0][ 7] = ff_avg_h264_qpel16_mc31_neon;
|
yading@10
|
145 c->avg_h264_qpel_pixels_tab[0][ 8] = ff_avg_h264_qpel16_mc02_neon;
|
yading@10
|
146 c->avg_h264_qpel_pixels_tab[0][ 9] = ff_avg_h264_qpel16_mc12_neon;
|
yading@10
|
147 c->avg_h264_qpel_pixels_tab[0][10] = ff_avg_h264_qpel16_mc22_neon;
|
yading@10
|
148 c->avg_h264_qpel_pixels_tab[0][11] = ff_avg_h264_qpel16_mc32_neon;
|
yading@10
|
149 c->avg_h264_qpel_pixels_tab[0][12] = ff_avg_h264_qpel16_mc03_neon;
|
yading@10
|
150 c->avg_h264_qpel_pixels_tab[0][13] = ff_avg_h264_qpel16_mc13_neon;
|
yading@10
|
151 c->avg_h264_qpel_pixels_tab[0][14] = ff_avg_h264_qpel16_mc23_neon;
|
yading@10
|
152 c->avg_h264_qpel_pixels_tab[0][15] = ff_avg_h264_qpel16_mc33_neon;
|
yading@10
|
153
|
yading@10
|
154 c->avg_h264_qpel_pixels_tab[1][ 0] = ff_avg_h264_qpel8_mc00_neon;
|
yading@10
|
155 c->avg_h264_qpel_pixels_tab[1][ 1] = ff_avg_h264_qpel8_mc10_neon;
|
yading@10
|
156 c->avg_h264_qpel_pixels_tab[1][ 2] = ff_avg_h264_qpel8_mc20_neon;
|
yading@10
|
157 c->avg_h264_qpel_pixels_tab[1][ 3] = ff_avg_h264_qpel8_mc30_neon;
|
yading@10
|
158 c->avg_h264_qpel_pixels_tab[1][ 4] = ff_avg_h264_qpel8_mc01_neon;
|
yading@10
|
159 c->avg_h264_qpel_pixels_tab[1][ 5] = ff_avg_h264_qpel8_mc11_neon;
|
yading@10
|
160 c->avg_h264_qpel_pixels_tab[1][ 6] = ff_avg_h264_qpel8_mc21_neon;
|
yading@10
|
161 c->avg_h264_qpel_pixels_tab[1][ 7] = ff_avg_h264_qpel8_mc31_neon;
|
yading@10
|
162 c->avg_h264_qpel_pixels_tab[1][ 8] = ff_avg_h264_qpel8_mc02_neon;
|
yading@10
|
163 c->avg_h264_qpel_pixels_tab[1][ 9] = ff_avg_h264_qpel8_mc12_neon;
|
yading@10
|
164 c->avg_h264_qpel_pixels_tab[1][10] = ff_avg_h264_qpel8_mc22_neon;
|
yading@10
|
165 c->avg_h264_qpel_pixels_tab[1][11] = ff_avg_h264_qpel8_mc32_neon;
|
yading@10
|
166 c->avg_h264_qpel_pixels_tab[1][12] = ff_avg_h264_qpel8_mc03_neon;
|
yading@10
|
167 c->avg_h264_qpel_pixels_tab[1][13] = ff_avg_h264_qpel8_mc13_neon;
|
yading@10
|
168 c->avg_h264_qpel_pixels_tab[1][14] = ff_avg_h264_qpel8_mc23_neon;
|
yading@10
|
169 c->avg_h264_qpel_pixels_tab[1][15] = ff_avg_h264_qpel8_mc33_neon;
|
yading@10
|
170 }
|
yading@10
|
171 }
|