annotate faanidct.c @ 8881:f8042554d4c8 libavcodec

Add xvmc_internal.h that contains all internal xvmc function declarations. This allows getting rid of a bunch of ugly forward declarations.
author diego
date Sat, 14 Feb 2009 20:37:14 +0000
parents b72f6c4cee12
children
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
6407
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
1 /*
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
2 * Floating point AAN IDCT
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
3 * Copyright (c) 2008 Michael Niedermayer <michaelni@gmx.at>
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
4 *
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
5 * This file is part of FFmpeg.
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
6 *
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
7 * FFmpeg is free software; you can redistribute it and/or
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
8 * modify it under the terms of the GNU Lesser General Public
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
9 * License as published by the Free Software Foundation; either
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
10 * version 2.1 of the License, or (at your option) any later version.
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
11 *
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
12 * FFmpeg is distributed in the hope that it will be useful,
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
15 * Lesser General Public License for more details.
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
16 *
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
17 * You should have received a copy of the GNU Lesser General Public
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
18 * License along with FFmpeg; if not, write to the Free Software
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
20 */
6421
53308ac1a5cc better include
michael
parents: 6418
diff changeset
21 #include "faanidct.h"
6407
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
22
6418
eb740c8e9212 Add explanatory comment to '#define FLOAT float'.
diego
parents: 6415
diff changeset
23 /* To allow switching to double. */
6407
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
24 #define FLOAT float
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
25
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
26 #define B0 1.0000000000000000000000
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
27 #define B1 1.3870398453221474618216 // cos(pi*1/16)sqrt(2)
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
28 #define B2 1.3065629648763765278566 // cos(pi*2/16)sqrt(2)
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
29 #define B3 1.1758756024193587169745 // cos(pi*3/16)sqrt(2)
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
30 #define B4 1.0000000000000000000000 // cos(pi*4/16)sqrt(2)
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
31 #define B5 0.7856949583871021812779 // cos(pi*5/16)sqrt(2)
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
32 #define B6 0.5411961001461969843997 // cos(pi*6/16)sqrt(2)
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
33 #define B7 0.2758993792829430123360 // cos(pi*7/16)sqrt(2)
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
34
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
35 #define A4 0.70710678118654752438 // cos(pi*4/16)
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
36 #define A2 0.92387953251128675613 // cos(pi*2/16)
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
37
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
38 static const FLOAT prescale[64]={
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
39 B0*B0/8, B0*B1/8, B0*B2/8, B0*B3/8, B0*B4/8, B0*B5/8, B0*B6/8, B0*B7/8,
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
40 B1*B0/8, B1*B1/8, B1*B2/8, B1*B3/8, B1*B4/8, B1*B5/8, B1*B6/8, B1*B7/8,
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
41 B2*B0/8, B2*B1/8, B2*B2/8, B2*B3/8, B2*B4/8, B2*B5/8, B2*B6/8, B2*B7/8,
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
42 B3*B0/8, B3*B1/8, B3*B2/8, B3*B3/8, B3*B4/8, B3*B5/8, B3*B6/8, B3*B7/8,
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
43 B4*B0/8, B4*B1/8, B4*B2/8, B4*B3/8, B4*B4/8, B4*B5/8, B4*B6/8, B4*B7/8,
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
44 B5*B0/8, B5*B1/8, B5*B2/8, B5*B3/8, B5*B4/8, B5*B5/8, B5*B6/8, B5*B7/8,
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
45 B6*B0/8, B6*B1/8, B6*B2/8, B6*B3/8, B6*B4/8, B6*B5/8, B6*B6/8, B6*B7/8,
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
46 B7*B0/8, B7*B1/8, B7*B2/8, B7*B3/8, B7*B4/8, B7*B5/8, B7*B6/8, B7*B7/8,
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
47 };
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
48
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
49 static inline void p8idct(DCTELEM data[64], FLOAT temp[64], uint8_t *dest, int stride, int x, int y, int type){
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
50 int i;
6471
b72f6c4cee12 Fix unused variable warnings.
diego
parents: 6421
diff changeset
51 FLOAT av_unused tmp0;
6407
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
52 FLOAT s04, d04, s17, d17, s26, d26, s53, d53;
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
53 FLOAT os07, os16, os25, os34;
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
54 FLOAT od07, od16, od25, od34;
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
55
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
56 for(i=0; i<y*8; i+=y){
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
57 s17= temp[1*x + i] + temp[7*x + i];
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
58 d17= temp[1*x + i] - temp[7*x + i];
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
59 s53= temp[5*x + i] + temp[3*x + i];
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
60 d53= temp[5*x + i] - temp[3*x + i];
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
61
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
62 od07= s17 + s53;
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
63 od25= (s17 - s53)*(2*A4);
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
64
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
65 #if 0 //these 2 are equivalent
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
66 tmp0= (d17 + d53)*(2*A2);
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
67 od34= d17*( 2*B6) - tmp0;
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
68 od16= d53*(-2*B2) + tmp0;
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
69 #else
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
70 od34= d17*(2*(B6-A2)) - d53*(2*A2);
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
71 od16= d53*(2*(A2-B2)) + d17*(2*A2);
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
72 #endif
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
73
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
74 od16 -= od07;
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
75 od25 -= od16;
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
76 od34 += od25;
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
77
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
78 s26 = temp[2*x + i] + temp[6*x + i];
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
79 d26 = temp[2*x + i] - temp[6*x + i];
6415
2eb1b18d6282 Remove another temporary variable with which gcc has problems.
michael
parents: 6407
diff changeset
80 d26*= 2*A4;
2eb1b18d6282 Remove another temporary variable with which gcc has problems.
michael
parents: 6407
diff changeset
81 d26-= s26;
6407
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
82
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
83 s04= temp[0*x + i] + temp[4*x + i];
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
84 d04= temp[0*x + i] - temp[4*x + i];
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
85
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
86 os07= s04 + s26;
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
87 os34= s04 - s26;
6415
2eb1b18d6282 Remove another temporary variable with which gcc has problems.
michael
parents: 6407
diff changeset
88 os16= d04 + d26;
2eb1b18d6282 Remove another temporary variable with which gcc has problems.
michael
parents: 6407
diff changeset
89 os25= d04 - d26;
6407
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
90
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
91 if(type==0){
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
92 temp[0*x + i]= os07 + od07;
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
93 temp[7*x + i]= os07 - od07;
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
94 temp[1*x + i]= os16 + od16;
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
95 temp[6*x + i]= os16 - od16;
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
96 temp[2*x + i]= os25 + od25;
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
97 temp[5*x + i]= os25 - od25;
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
98 temp[3*x + i]= os34 - od34;
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
99 temp[4*x + i]= os34 + od34;
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
100 }else if(type==1){
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
101 data[0*x + i]= lrintf(os07 + od07);
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
102 data[7*x + i]= lrintf(os07 - od07);
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
103 data[1*x + i]= lrintf(os16 + od16);
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
104 data[6*x + i]= lrintf(os16 - od16);
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
105 data[2*x + i]= lrintf(os25 + od25);
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
106 data[5*x + i]= lrintf(os25 - od25);
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
107 data[3*x + i]= lrintf(os34 - od34);
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
108 data[4*x + i]= lrintf(os34 + od34);
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
109 }else if(type==2){
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
110 dest[0*stride + i]= av_clip_uint8(((int)dest[0*stride + i]) + lrintf(os07 + od07));
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
111 dest[7*stride + i]= av_clip_uint8(((int)dest[7*stride + i]) + lrintf(os07 - od07));
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
112 dest[1*stride + i]= av_clip_uint8(((int)dest[1*stride + i]) + lrintf(os16 + od16));
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
113 dest[6*stride + i]= av_clip_uint8(((int)dest[6*stride + i]) + lrintf(os16 - od16));
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
114 dest[2*stride + i]= av_clip_uint8(((int)dest[2*stride + i]) + lrintf(os25 + od25));
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
115 dest[5*stride + i]= av_clip_uint8(((int)dest[5*stride + i]) + lrintf(os25 - od25));
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
116 dest[3*stride + i]= av_clip_uint8(((int)dest[3*stride + i]) + lrintf(os34 - od34));
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
117 dest[4*stride + i]= av_clip_uint8(((int)dest[4*stride + i]) + lrintf(os34 + od34));
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
118 }else{
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
119 dest[0*stride + i]= av_clip_uint8(lrintf(os07 + od07));
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
120 dest[7*stride + i]= av_clip_uint8(lrintf(os07 - od07));
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
121 dest[1*stride + i]= av_clip_uint8(lrintf(os16 + od16));
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
122 dest[6*stride + i]= av_clip_uint8(lrintf(os16 - od16));
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
123 dest[2*stride + i]= av_clip_uint8(lrintf(os25 + od25));
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
124 dest[5*stride + i]= av_clip_uint8(lrintf(os25 - od25));
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
125 dest[3*stride + i]= av_clip_uint8(lrintf(os34 - od34));
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
126 dest[4*stride + i]= av_clip_uint8(lrintf(os34 + od34));
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
127 }
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
128 }
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
129 }
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
130
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
131 void ff_faanidct(DCTELEM block[64]){
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
132 FLOAT temp[64];
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
133 int i;
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
134
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
135 emms_c();
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
136
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
137 for(i=0; i<64; i++)
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
138 temp[i] = block[i] * prescale[i];
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
139
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
140 p8idct(block, temp, NULL, 0, 1, 8, 0);
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
141 p8idct(block, temp, NULL, 0, 8, 1, 1);
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
142 }
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
143
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
144 void ff_faanidct_add(uint8_t *dest, int line_size, DCTELEM block[64]){
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
145 FLOAT temp[64];
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
146 int i;
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
147
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
148 emms_c();
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
149
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
150 for(i=0; i<64; i++)
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
151 temp[i] = block[i] * prescale[i];
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
152
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
153 p8idct(block, temp, NULL, 0, 1, 8, 0);
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
154 p8idct(NULL , temp, dest, line_size, 8, 1, 2);
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
155 }
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
156
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
157 void ff_faanidct_put(uint8_t *dest, int line_size, DCTELEM block[64]){
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
158 FLOAT temp[64];
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
159 int i;
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
160
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
161 emms_c();
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
162
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
163 for(i=0; i<64; i++)
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
164 temp[i] = block[i] * prescale[i];
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
165
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
166 p8idct(block, temp, NULL, 0, 1, 8, 0);
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
167 p8idct(NULL , temp, dest, line_size, 8, 1, 3);
7f9e2b5893fc floating point AAN IDCT
michael
parents:
diff changeset
168 }