annotate ppc/idct_altivec.c @ 7422:e29ac68d3152 libavcodec

AAC tables and data that have been OKed by Michael
author superdump
date Sun, 27 Jul 2008 08:21:48 +0000
parents 266d4949aa15
children 1615d6b75ada
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
1 /*
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
2 * Copyright (c) 2001 Michel Lespinasse
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
3 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3036
diff changeset
4 * This file is part of FFmpeg.
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3036
diff changeset
5 *
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3036
diff changeset
6 * FFmpeg is free software; you can redistribute it and/or
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
7 * modify it under the terms of the GNU Lesser General Public
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
8 * 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: 3036
diff changeset
9 * version 2.1 of the License, or (at your option) any later version.
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
10 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3036
diff changeset
11 * FFmpeg is distributed in the hope that it will be useful,
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
14 * Lesser General Public License for more details.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
15 *
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
16 * 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: 3036
diff changeset
17 * License along with FFmpeg; if not, write to the Free Software
3036
0b546eab515d Update licensing information: The FSF changed postal address.
diego
parents: 2979
diff changeset
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
19 */
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
20
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
21 /*
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
22 * NOTE: This code is based on GPL code from the libmpeg2 project. The
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
23 * author, Michel Lespinasses, has given explicit permission to release
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
24 * under LGPL as part of ffmpeg.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
25 */
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
26
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
27 /*
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
28 * FFMpeg integration by Dieter Shirley
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
29 *
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
30 * This file is a direct copy of the altivec idct module from the libmpeg2
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
31 * project. I've deleted all of the libmpeg2 specific code, renamed the functions and
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
32 * re-ordered the function parameters. The only change to the IDCT function
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
33 * itself was to factor out the partial transposition, and to perform a full
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
34 * transpose at the end of the function.
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
35 */
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
36
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
37
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
38 #include <stdlib.h> /* malloc(), free() */
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
39 #include <string.h>
6763
f7cbb7733146 Use full path for #includes from another directory.
diego
parents: 6105
diff changeset
40 #include "libavcodec/dsputil.h"
1277
f3152eb76f1a altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents: 1064
diff changeset
41
f3152eb76f1a altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents: 1064
diff changeset
42 #include "gcc_fixes.h"
f3152eb76f1a altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents: 1064
diff changeset
43
6105
33674fb857b5 Change some files to only include the necessary headers.
diego
parents: 5746
diff changeset
44 #include "dsputil_ppc.h"
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
45
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
46 #define vector_s16_t vector signed short
5746
55ed6dc5d476 Remove const vector macro indirection that is useless and obfuscating
diego
parents: 5215
diff changeset
47 #define const_vector_s16_t const vector signed short
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
48 #define vector_u16_t vector unsigned short
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
49 #define vector_s8_t vector signed char
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
50 #define vector_u8_t vector unsigned char
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
51 #define vector_s32_t vector signed int
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
52 #define vector_u32_t vector unsigned int
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
53
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
54 #define IDCT_HALF \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
55 /* 1st stage */ \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
56 t1 = vec_mradds (a1, vx7, vx1 ); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
57 t8 = vec_mradds (a1, vx1, vec_subs (zero, vx7)); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
58 t7 = vec_mradds (a2, vx5, vx3); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
59 t3 = vec_mradds (ma2, vx3, vx5); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
60 \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
61 /* 2nd stage */ \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
62 t5 = vec_adds (vx0, vx4); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
63 t0 = vec_subs (vx0, vx4); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
64 t2 = vec_mradds (a0, vx6, vx2); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
65 t4 = vec_mradds (a0, vx2, vec_subs (zero, vx6)); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
66 t6 = vec_adds (t8, t3); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
67 t3 = vec_subs (t8, t3); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
68 t8 = vec_subs (t1, t7); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
69 t1 = vec_adds (t1, t7); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
70 \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
71 /* 3rd stage */ \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
72 t7 = vec_adds (t5, t2); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
73 t2 = vec_subs (t5, t2); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
74 t5 = vec_adds (t0, t4); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
75 t0 = vec_subs (t0, t4); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
76 t4 = vec_subs (t8, t3); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
77 t3 = vec_adds (t8, t3); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
78 \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
79 /* 4th stage */ \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
80 vy0 = vec_adds (t7, t1); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
81 vy7 = vec_subs (t7, t1); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
82 vy1 = vec_mradds (c4, t3, t5); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
83 vy6 = vec_mradds (mc4, t3, t5); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
84 vy2 = vec_mradds (c4, t4, t0); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
85 vy5 = vec_mradds (mc4, t4, t0); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
86 vy3 = vec_adds (t2, t6); \
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
87 vy4 = vec_subs (t2, t6);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
88
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1839
diff changeset
89
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
90 #define IDCT \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
91 vector_s16_t vx0, vx1, vx2, vx3, vx4, vx5, vx6, vx7; \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
92 vector_s16_t vy0, vy1, vy2, vy3, vy4, vy5, vy6, vy7; \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
93 vector_s16_t a0, a1, a2, ma2, c4, mc4, zero, bias; \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
94 vector_s16_t t0, t1, t2, t3, t4, t5, t6, t7, t8; \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
95 vector_u16_t shift; \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
96 \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
97 c4 = vec_splat (constants[0], 0); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
98 a0 = vec_splat (constants[0], 1); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
99 a1 = vec_splat (constants[0], 2); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
100 a2 = vec_splat (constants[0], 3); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
101 mc4 = vec_splat (constants[0], 4); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
102 ma2 = vec_splat (constants[0], 5); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
103 bias = (vector_s16_t)vec_splat ((vector_s32_t)constants[0], 3); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
104 \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
105 zero = vec_splat_s16 (0); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
106 shift = vec_splat_u16 (4); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
107 \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
108 vx0 = vec_mradds (vec_sl (block[0], shift), constants[1], zero); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
109 vx1 = vec_mradds (vec_sl (block[1], shift), constants[2], zero); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
110 vx2 = vec_mradds (vec_sl (block[2], shift), constants[3], zero); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
111 vx3 = vec_mradds (vec_sl (block[3], shift), constants[4], zero); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
112 vx4 = vec_mradds (vec_sl (block[4], shift), constants[1], zero); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
113 vx5 = vec_mradds (vec_sl (block[5], shift), constants[4], zero); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
114 vx6 = vec_mradds (vec_sl (block[6], shift), constants[3], zero); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
115 vx7 = vec_mradds (vec_sl (block[7], shift), constants[2], zero); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
116 \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
117 IDCT_HALF \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
118 \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
119 vx0 = vec_mergeh (vy0, vy4); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
120 vx1 = vec_mergel (vy0, vy4); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
121 vx2 = vec_mergeh (vy1, vy5); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
122 vx3 = vec_mergel (vy1, vy5); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
123 vx4 = vec_mergeh (vy2, vy6); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
124 vx5 = vec_mergel (vy2, vy6); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
125 vx6 = vec_mergeh (vy3, vy7); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
126 vx7 = vec_mergel (vy3, vy7); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
127 \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
128 vy0 = vec_mergeh (vx0, vx4); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
129 vy1 = vec_mergel (vx0, vx4); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
130 vy2 = vec_mergeh (vx1, vx5); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
131 vy3 = vec_mergel (vx1, vx5); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
132 vy4 = vec_mergeh (vx2, vx6); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
133 vy5 = vec_mergel (vx2, vx6); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
134 vy6 = vec_mergeh (vx3, vx7); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
135 vy7 = vec_mergel (vx3, vx7); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
136 \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
137 vx0 = vec_adds (vec_mergeh (vy0, vy4), bias); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
138 vx1 = vec_mergel (vy0, vy4); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
139 vx2 = vec_mergeh (vy1, vy5); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
140 vx3 = vec_mergel (vy1, vy5); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
141 vx4 = vec_mergeh (vy2, vy6); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
142 vx5 = vec_mergel (vy2, vy6); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
143 vx6 = vec_mergeh (vy3, vy7); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
144 vx7 = vec_mergel (vy3, vy7); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
145 \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
146 IDCT_HALF \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
147 \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
148 shift = vec_splat_u16 (6); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
149 vx0 = vec_sra (vy0, shift); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
150 vx1 = vec_sra (vy1, shift); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
151 vx2 = vec_sra (vy2, shift); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
152 vx3 = vec_sra (vy3, shift); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
153 vx4 = vec_sra (vy4, shift); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
154 vx5 = vec_sra (vy5, shift); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
155 vx6 = vec_sra (vy6, shift); \
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
156 vx7 = vec_sra (vy7, shift);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
157
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1015
diff changeset
158
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1352
diff changeset
159 static const_vector_s16_t constants[5] = {
7373
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7333
diff changeset
160 {23170, 13573, 6518, 21895, -23170, -21895, 32, 31},
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7333
diff changeset
161 {16384, 22725, 21407, 19266, 16384, 19266, 21407, 22725},
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7333
diff changeset
162 {22725, 31521, 29692, 26722, 22725, 26722, 29692, 31521},
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7333
diff changeset
163 {21407, 29692, 27969, 25172, 21407, 25172, 27969, 29692},
266d4949aa15 Remove AltiVec vector declaration compiler compatibility macros.
diego
parents: 7333
diff changeset
164 {19266, 26722, 25172, 22654, 19266, 22654, 25172, 26722}
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
165 };
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
166
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
167 void idct_put_altivec(uint8_t* dest, int stride, vector_s16_t* block)
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
168 {
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1277
diff changeset
169 POWERPC_PERF_DECLARE(altivec_idct_put_num, 1);
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
170 vector_u8_t tmp;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
171
4521
891590781d9e rename POWERPC_PERFORMANCE_REPORT to CONFIG_POWERPC_PERF
mru
parents: 3973
diff changeset
172 #ifdef CONFIG_POWERPC_PERF
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1277
diff changeset
173 POWERPC_PERF_START_COUNT(altivec_idct_put_num, 1);
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1352
diff changeset
174 #endif
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
175 IDCT
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
176
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
177 #define COPY(dest,src) \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
178 tmp = vec_packsu (src, src); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
179 vec_ste ((vector_u32_t)tmp, 0, (unsigned int *)dest); \
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
180 vec_ste ((vector_u32_t)tmp, 4, (unsigned int *)dest);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
181
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
182 COPY (dest, vx0) dest += stride;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
183 COPY (dest, vx1) dest += stride;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
184 COPY (dest, vx2) dest += stride;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
185 COPY (dest, vx3) dest += stride;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
186 COPY (dest, vx4) dest += stride;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
187 COPY (dest, vx5) dest += stride;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
188 COPY (dest, vx6) dest += stride;
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
189 COPY (dest, vx7)
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 828
diff changeset
190
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1277
diff changeset
191 POWERPC_PERF_STOP_COUNT(altivec_idct_put_num, 1);
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
192 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
193
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
194 void idct_add_altivec(uint8_t* dest, int stride, vector_s16_t* block)
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
195 {
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1277
diff changeset
196 POWERPC_PERF_DECLARE(altivec_idct_add_num, 1);
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
197 vector_u8_t tmp;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
198 vector_s16_t tmp2, tmp3;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
199 vector_u8_t perm0;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
200 vector_u8_t perm1;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
201 vector_u8_t p0, p1, p;
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
202
4521
891590781d9e rename POWERPC_PERFORMANCE_REPORT to CONFIG_POWERPC_PERF
mru
parents: 3973
diff changeset
203 #ifdef CONFIG_POWERPC_PERF
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1277
diff changeset
204 POWERPC_PERF_START_COUNT(altivec_idct_add_num, 1);
1839
b370288f004d Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents: 1352
diff changeset
205 #endif
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 828
diff changeset
206
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
207 IDCT
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
208
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
209 p0 = vec_lvsl (0, dest);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
210 p1 = vec_lvsl (stride, dest);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
211 p = vec_splat_u8 (-1);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
212 perm0 = vec_mergeh (p, p0);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
213 perm1 = vec_mergeh (p, p1);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
214
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
215 #define ADD(dest,src,perm) \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
216 /* *(uint64_t *)&tmp = *(uint64_t *)dest; */ \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
217 tmp = vec_ld (0, dest); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
218 tmp2 = (vector_s16_t)vec_perm (tmp, (vector_u8_t)zero, perm); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
219 tmp3 = vec_adds (tmp2, src); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
220 tmp = vec_packsu (tmp3, tmp3); \
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
221 vec_ste ((vector_u32_t)tmp, 0, (unsigned int *)dest); \
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
222 vec_ste ((vector_u32_t)tmp, 4, (unsigned int *)dest);
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
223
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
224 ADD (dest, vx0, perm0) dest += stride;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
225 ADD (dest, vx1, perm1) dest += stride;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
226 ADD (dest, vx2, perm0) dest += stride;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
227 ADD (dest, vx3, perm1) dest += stride;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
228 ADD (dest, vx4, perm0) dest += stride;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
229 ADD (dest, vx5, perm1) dest += stride;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
230 ADD (dest, vx6, perm0) dest += stride;
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
231 ADD (dest, vx7, perm1)
1009
3b7cc8e4b83f AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents: 828
diff changeset
232
1352
e8ff4783f188 1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents: 1277
diff changeset
233 POWERPC_PERF_STOP_COUNT(altivec_idct_add_num, 1);
828
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
234 }
ace3ccd18dd2 Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff changeset
235