yading@10
|
1 /*
|
yading@10
|
2 * RV30/40 decoder common dsp functions
|
yading@10
|
3 * Copyright (c) 2007 Mike Melanson, Konstantin Shishkov
|
yading@10
|
4 * Copyright (c) 2011 Janne Grunau
|
yading@10
|
5 *
|
yading@10
|
6 * This file is part of Libav.
|
yading@10
|
7 *
|
yading@10
|
8 * Libav is free software; you can redistribute it and/or
|
yading@10
|
9 * modify it under the terms of the GNU Lesser General Public
|
yading@10
|
10 * License as published by the Free Software Foundation; either
|
yading@10
|
11 * version 2.1 of the License, or (at your option) any later version.
|
yading@10
|
12 *
|
yading@10
|
13 * Libav is distributed in the hope that it will be useful,
|
yading@10
|
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
yading@10
|
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
yading@10
|
16 * Lesser General Public License for more details.
|
yading@10
|
17 *
|
yading@10
|
18 * You should have received a copy of the GNU Lesser General Public
|
yading@10
|
19 * License along with Libav; if not, write to the Free Software
|
yading@10
|
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
yading@10
|
21 */
|
yading@10
|
22
|
yading@10
|
23 /**
|
yading@10
|
24 * @file
|
yading@10
|
25 * RV30/40 decoder common dsp functions
|
yading@10
|
26 */
|
yading@10
|
27
|
yading@10
|
28 #include "rv34dsp.h"
|
yading@10
|
29 #include "libavutil/common.h"
|
yading@10
|
30
|
yading@10
|
31 /**
|
yading@10
|
32 * @name RV30/40 inverse transform functions
|
yading@10
|
33 * @{
|
yading@10
|
34 */
|
yading@10
|
35
|
yading@10
|
36 static av_always_inline void rv34_row_transform(int temp[16], int16_t *block)
|
yading@10
|
37 {
|
yading@10
|
38 int i;
|
yading@10
|
39
|
yading@10
|
40 for(i = 0; i < 4; i++){
|
yading@10
|
41 const int z0 = 13*(block[i+4*0] + block[i+4*2]);
|
yading@10
|
42 const int z1 = 13*(block[i+4*0] - block[i+4*2]);
|
yading@10
|
43 const int z2 = 7* block[i+4*1] - 17*block[i+4*3];
|
yading@10
|
44 const int z3 = 17* block[i+4*1] + 7*block[i+4*3];
|
yading@10
|
45
|
yading@10
|
46 temp[4*i+0] = z0 + z3;
|
yading@10
|
47 temp[4*i+1] = z1 + z2;
|
yading@10
|
48 temp[4*i+2] = z1 - z2;
|
yading@10
|
49 temp[4*i+3] = z0 - z3;
|
yading@10
|
50 }
|
yading@10
|
51 }
|
yading@10
|
52
|
yading@10
|
53 /**
|
yading@10
|
54 * Real Video 3.0/4.0 inverse transform + sample reconstruction
|
yading@10
|
55 * Code is almost the same as in SVQ3, only scaling is different.
|
yading@10
|
56 */
|
yading@10
|
57 static void rv34_idct_add_c(uint8_t *dst, ptrdiff_t stride, int16_t *block){
|
yading@10
|
58 int temp[16];
|
yading@10
|
59 int i;
|
yading@10
|
60
|
yading@10
|
61 rv34_row_transform(temp, block);
|
yading@10
|
62 memset(block, 0, 16*sizeof(int16_t));
|
yading@10
|
63
|
yading@10
|
64 for(i = 0; i < 4; i++){
|
yading@10
|
65 const int z0 = 13*(temp[4*0+i] + temp[4*2+i]) + 0x200;
|
yading@10
|
66 const int z1 = 13*(temp[4*0+i] - temp[4*2+i]) + 0x200;
|
yading@10
|
67 const int z2 = 7* temp[4*1+i] - 17*temp[4*3+i];
|
yading@10
|
68 const int z3 = 17* temp[4*1+i] + 7*temp[4*3+i];
|
yading@10
|
69
|
yading@10
|
70 dst[0] = av_clip_uint8( dst[0] + ( (z0 + z3) >> 10 ) );
|
yading@10
|
71 dst[1] = av_clip_uint8( dst[1] + ( (z1 + z2) >> 10 ) );
|
yading@10
|
72 dst[2] = av_clip_uint8( dst[2] + ( (z1 - z2) >> 10 ) );
|
yading@10
|
73 dst[3] = av_clip_uint8( dst[3] + ( (z0 - z3) >> 10 ) );
|
yading@10
|
74
|
yading@10
|
75 dst += stride;
|
yading@10
|
76 }
|
yading@10
|
77 }
|
yading@10
|
78
|
yading@10
|
79 /**
|
yading@10
|
80 * RealVideo 3.0/4.0 inverse transform for DC block
|
yading@10
|
81 *
|
yading@10
|
82 * Code is almost the same as rv34_inv_transform()
|
yading@10
|
83 * but final coefficients are multiplied by 1.5 and have no rounding.
|
yading@10
|
84 */
|
yading@10
|
85 static void rv34_inv_transform_noround_c(int16_t *block){
|
yading@10
|
86 int temp[16];
|
yading@10
|
87 int i;
|
yading@10
|
88
|
yading@10
|
89 rv34_row_transform(temp, block);
|
yading@10
|
90
|
yading@10
|
91 for(i = 0; i < 4; i++){
|
yading@10
|
92 const int z0 = 39*(temp[4*0+i] + temp[4*2+i]);
|
yading@10
|
93 const int z1 = 39*(temp[4*0+i] - temp[4*2+i]);
|
yading@10
|
94 const int z2 = 21* temp[4*1+i] - 51*temp[4*3+i];
|
yading@10
|
95 const int z3 = 51* temp[4*1+i] + 21*temp[4*3+i];
|
yading@10
|
96
|
yading@10
|
97 block[i*4+0] = (z0 + z3) >> 11;
|
yading@10
|
98 block[i*4+1] = (z1 + z2) >> 11;
|
yading@10
|
99 block[i*4+2] = (z1 - z2) >> 11;
|
yading@10
|
100 block[i*4+3] = (z0 - z3) >> 11;
|
yading@10
|
101 }
|
yading@10
|
102 }
|
yading@10
|
103
|
yading@10
|
104 static void rv34_idct_dc_add_c(uint8_t *dst, ptrdiff_t stride, int dc)
|
yading@10
|
105 {
|
yading@10
|
106 int i, j;
|
yading@10
|
107
|
yading@10
|
108 dc = (13*13*dc + 0x200) >> 10;
|
yading@10
|
109 for (i = 0; i < 4; i++)
|
yading@10
|
110 {
|
yading@10
|
111 for (j = 0; j < 4; j++)
|
yading@10
|
112 dst[j] = av_clip_uint8( dst[j] + dc );
|
yading@10
|
113
|
yading@10
|
114 dst += stride;
|
yading@10
|
115 }
|
yading@10
|
116 }
|
yading@10
|
117
|
yading@10
|
118 static void rv34_inv_transform_dc_noround_c(int16_t *block)
|
yading@10
|
119 {
|
yading@10
|
120 int16_t dc = (13 * 13 * 3 * block[0]) >> 11;
|
yading@10
|
121 int i, j;
|
yading@10
|
122
|
yading@10
|
123 for (i = 0; i < 4; i++, block += 4)
|
yading@10
|
124 for (j = 0; j < 4; j++)
|
yading@10
|
125 block[j] = dc;
|
yading@10
|
126 }
|
yading@10
|
127
|
yading@10
|
128 /** @} */ // transform
|
yading@10
|
129
|
yading@10
|
130
|
yading@10
|
131 av_cold void ff_rv34dsp_init(RV34DSPContext *c)
|
yading@10
|
132 {
|
yading@10
|
133 c->rv34_inv_transform = rv34_inv_transform_noround_c;
|
yading@10
|
134 c->rv34_inv_transform_dc = rv34_inv_transform_dc_noround_c;
|
yading@10
|
135
|
yading@10
|
136 c->rv34_idct_add = rv34_idct_add_c;
|
yading@10
|
137 c->rv34_idct_dc_add = rv34_idct_dc_add_c;
|
yading@10
|
138
|
yading@10
|
139 if (ARCH_ARM)
|
yading@10
|
140 ff_rv34dsp_init_arm(c);
|
yading@10
|
141 if (ARCH_X86)
|
yading@10
|
142 ff_rv34dsp_init_x86(c);
|
yading@10
|
143 }
|