]> git.sesse.net Git - x264/blob - common/aarch64/predict-c.c
aarch64: NEON asm for 8x16c intra prediction
[x264] / common / aarch64 / predict-c.c
1 /*****************************************************************************
2  * predict.c: aarch64 intra prediction
3  *****************************************************************************
4  * Copyright (C) 2009-2014 x264 project
5  *
6  * Authors: David Conrad <lessen42@gmail.com>
7  *          Janne Grunau <janne-x264@jannau.net>
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of the GNU General Public License as published by
11  * the Free Software Foundation; either version 2 of the License, or
12  * (at your option) any later version.
13  *
14  * This program is distributed in the hope that it will be useful,
15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
17  * GNU General Public License for more details.
18  *
19  * You should have received a copy of the GNU General Public License
20  * along with this program; if not, write to the Free Software
21  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02111, USA.
22  *
23  * This program is also available under a commercial proprietary license.
24  * For more information, contact us at licensing@x264.com.
25  *****************************************************************************/
26
27 #include "common/common.h"
28 #include "predict.h"
29 #include "pixel.h"
30
31 void x264_predict_4x4_dc_top_neon( uint8_t *src );
32 void x264_predict_4x4_ddr_neon( uint8_t *src );
33 void x264_predict_4x4_ddl_neon( uint8_t *src );
34
35 void x264_predict_8x8c_dc_top_neon( uint8_t *src );
36 void x264_predict_8x8c_dc_left_neon( uint8_t *src );
37 void x264_predict_8x8c_p_neon( uint8_t *src );
38
39 void x264_predict_8x16c_dc_left_neon( uint8_t *src );
40 void x264_predict_8x16c_dc_top_neon( uint8_t *src );
41 void x264_predict_8x16c_p_neon( uint8_t *src );
42
43 void x264_predict_8x8_ddl_neon( uint8_t *src, uint8_t edge[36] );
44 void x264_predict_8x8_ddr_neon( uint8_t *src, uint8_t edge[36] );
45 void x264_predict_8x8_vl_neon( uint8_t *src, uint8_t edge[36] );
46 void x264_predict_8x8_vr_neon( uint8_t *src, uint8_t edge[36] );
47 void x264_predict_8x8_hd_neon( uint8_t *src, uint8_t edge[36] );
48 void x264_predict_8x8_hu_neon( uint8_t *src, uint8_t edge[36] );
49
50 void x264_predict_16x16_dc_top_neon( uint8_t *src );
51 void x264_predict_16x16_dc_left_neon( uint8_t *src );
52 void x264_predict_16x16_p_neon( uint8_t *src );
53
54 void x264_predict_4x4_init_aarch64( int cpu, x264_predict_t pf[12] )
55 {
56 #if !HIGH_BIT_DEPTH
57     if (cpu&X264_CPU_ARMV8)
58     {
59         pf[I_PRED_4x4_H]   = x264_predict_4x4_h_aarch64;
60         pf[I_PRED_4x4_V]   = x264_predict_4x4_v_aarch64;
61     }
62
63     if (cpu&X264_CPU_NEON)
64     {
65         pf[I_PRED_4x4_DC]     = x264_predict_4x4_dc_neon;
66         pf[I_PRED_4x4_DC_TOP] = x264_predict_4x4_dc_top_neon;
67         pf[I_PRED_4x4_DDL]    = x264_predict_4x4_ddl_neon;
68         pf[I_PRED_4x4_DDR]    = x264_predict_4x4_ddr_neon;
69     }
70 #endif // !HIGH_BIT_DEPTH
71 }
72
73 void x264_predict_8x8c_init_aarch64( int cpu, x264_predict_t pf[7] )
74 {
75     if (!(cpu&X264_CPU_NEON))
76         return;
77
78 #if !HIGH_BIT_DEPTH
79     pf[I_PRED_CHROMA_DC]      = x264_predict_8x8c_dc_neon;
80     pf[I_PRED_CHROMA_DC_TOP]  = x264_predict_8x8c_dc_top_neon;
81     pf[I_PRED_CHROMA_DC_LEFT] = x264_predict_8x8c_dc_left_neon;
82     pf[I_PRED_CHROMA_H]       = x264_predict_8x8c_h_neon;
83     pf[I_PRED_CHROMA_V]       = x264_predict_8x8c_v_neon;
84     pf[I_PRED_CHROMA_P]       = x264_predict_8x8c_p_neon;
85 #endif // !HIGH_BIT_DEPTH
86 }
87
88
89 void x264_predict_8x16c_init_aarch64( int cpu, x264_predict_t pf[7] )
90 {
91     if (!(cpu&X264_CPU_NEON))
92         return;
93
94 #if !HIGH_BIT_DEPTH
95     pf[I_PRED_CHROMA_V ]     = x264_predict_8x16c_v_neon;
96     pf[I_PRED_CHROMA_H ]     = x264_predict_8x16c_h_neon;
97     pf[I_PRED_CHROMA_DC]     = x264_predict_8x16c_dc_neon;
98     pf[I_PRED_CHROMA_P ]     = x264_predict_8x16c_p_neon;
99     pf[I_PRED_CHROMA_DC_LEFT]= x264_predict_8x16c_dc_left_neon;
100     pf[I_PRED_CHROMA_DC_TOP ]= x264_predict_8x16c_dc_top_neon;
101 #endif // !HIGH_BIT_DEPTH
102 }
103
104 void x264_predict_8x8_init_aarch64( int cpu, x264_predict8x8_t pf[12], x264_predict_8x8_filter_t *predict_filter )
105 {
106     if (!(cpu&X264_CPU_NEON))
107         return;
108
109 #if !HIGH_BIT_DEPTH
110     pf[I_PRED_8x8_DDL] = x264_predict_8x8_ddl_neon;
111     pf[I_PRED_8x8_DDR] = x264_predict_8x8_ddr_neon;
112     pf[I_PRED_8x8_VL]  = x264_predict_8x8_vl_neon;
113     pf[I_PRED_8x8_VR]  = x264_predict_8x8_vr_neon;
114     pf[I_PRED_8x8_DC]  = x264_predict_8x8_dc_neon;
115     pf[I_PRED_8x8_H]   = x264_predict_8x8_h_neon;
116     pf[I_PRED_8x8_HD]  = x264_predict_8x8_hd_neon;
117     pf[I_PRED_8x8_HU]  = x264_predict_8x8_hu_neon;
118     pf[I_PRED_8x8_V]   = x264_predict_8x8_v_neon;
119 #endif // !HIGH_BIT_DEPTH
120 }
121
122 void x264_predict_16x16_init_aarch64( int cpu, x264_predict_t pf[7] )
123 {
124     if (!(cpu&X264_CPU_NEON))
125         return;
126
127 #if !HIGH_BIT_DEPTH
128     pf[I_PRED_16x16_DC ]    = x264_predict_16x16_dc_neon;
129     pf[I_PRED_16x16_DC_TOP] = x264_predict_16x16_dc_top_neon;
130     pf[I_PRED_16x16_DC_LEFT]= x264_predict_16x16_dc_left_neon;
131     pf[I_PRED_16x16_H ]     = x264_predict_16x16_h_neon;
132     pf[I_PRED_16x16_V ]     = x264_predict_16x16_v_neon;
133     pf[I_PRED_16x16_P ]     = x264_predict_16x16_p_neon;
134 #endif // !HIGH_BIT_DEPTH
135 }