Mercurial > libavcodec.hg
annotate ppc/idct_altivec.c @ 1775:3875b8b30399 libavcodec
changed includes
author | alex |
---|---|
date | Sat, 31 Jan 2004 22:58:32 +0000 |
parents | e8ff4783f188 |
children | b370288f004d |
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 * |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
4 * This library is free software; you can redistribute it and/or |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
5 * 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
|
6 * License as published by the Free Software Foundation; either |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
7 * version 2 of the License, or (at your option) any later version. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
8 * |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
9 * This library is distributed in the hope that it will be useful, |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
10 * 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
|
11 * 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
|
12 * Lesser General Public License for more details. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
13 * |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
14 * You should have received a copy of the GNU Lesser General Public |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
15 * License along with this library; if not, write to the Free Software |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
17 * |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
18 */ |
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 * 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
|
22 * 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
|
23 * under LGPL as part of ffmpeg. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
24 * |
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> |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
40 #include "../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 |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
828
diff
changeset
|
44 #include "dsputil_altivec.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 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
47 #define vector_u16_t vector unsigned short |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
48 #define vector_s8_t vector signed char |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
49 #define vector_u8_t vector unsigned char |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
50 #define vector_s32_t vector signed int |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
51 #define vector_u32_t vector unsigned int |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
52 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
53 #define IDCT_HALF \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
54 /* 1st stage */ \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
55 t1 = vec_mradds (a1, vx7, vx1 ); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
56 t8 = vec_mradds (a1, vx1, vec_subs (zero, vx7)); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
57 t7 = vec_mradds (a2, vx5, vx3); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
58 t3 = vec_mradds (ma2, vx3, vx5); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
59 \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
60 /* 2nd stage */ \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
61 t5 = vec_adds (vx0, vx4); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
62 t0 = vec_subs (vx0, vx4); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
63 t2 = vec_mradds (a0, vx6, vx2); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
64 t4 = vec_mradds (a0, vx2, vec_subs (zero, vx6)); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
65 t6 = vec_adds (t8, t3); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
66 t3 = vec_subs (t8, t3); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
67 t8 = vec_subs (t1, t7); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
68 t1 = vec_adds (t1, t7); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
69 \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
70 /* 3rd stage */ \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
71 t7 = vec_adds (t5, t2); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
72 t2 = vec_subs (t5, t2); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
73 t5 = vec_adds (t0, t4); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
74 t0 = vec_subs (t0, t4); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
75 t4 = vec_subs (t8, t3); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
76 t3 = vec_adds (t8, t3); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
77 \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
78 /* 4th stage */ \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
79 vy0 = vec_adds (t7, t1); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
80 vy7 = vec_subs (t7, t1); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
81 vy1 = vec_mradds (c4, t3, t5); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
82 vy6 = vec_mradds (mc4, t3, t5); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
83 vy2 = vec_mradds (c4, t4, t0); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
84 vy5 = vec_mradds (mc4, t4, t0); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
85 vy3 = vec_adds (t2, t6); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
86 vy4 = vec_subs (t2, t6); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
87 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
88 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
89 #define IDCT \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
90 vector_s16_t vx0, vx1, vx2, vx3, vx4, vx5, vx6, vx7; \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
91 vector_s16_t vy0, vy1, vy2, vy3, vy4, vy5, vy6, vy7; \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
92 vector_s16_t a0, a1, a2, ma2, c4, mc4, zero, bias; \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
93 vector_s16_t t0, t1, t2, t3, t4, t5, t6, t7, t8; \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
94 vector_u16_t shift; \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
95 \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
96 c4 = vec_splat (constants[0], 0); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
97 a0 = vec_splat (constants[0], 1); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
98 a1 = vec_splat (constants[0], 2); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
99 a2 = vec_splat (constants[0], 3); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
100 mc4 = vec_splat (constants[0], 4); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
101 ma2 = vec_splat (constants[0], 5); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
102 bias = (vector_s16_t)vec_splat ((vector_s32_t)constants[0], 3); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
103 \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
104 zero = vec_splat_s16 (0); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
105 shift = vec_splat_u16 (4); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
106 \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
107 vx0 = vec_mradds (vec_sl (block[0], shift), constants[1], zero); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
108 vx1 = vec_mradds (vec_sl (block[1], shift), constants[2], zero); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
109 vx2 = vec_mradds (vec_sl (block[2], shift), constants[3], zero); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
110 vx3 = vec_mradds (vec_sl (block[3], shift), constants[4], zero); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
111 vx4 = vec_mradds (vec_sl (block[4], shift), constants[1], zero); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
112 vx5 = vec_mradds (vec_sl (block[5], shift), constants[4], zero); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
113 vx6 = vec_mradds (vec_sl (block[6], shift), constants[3], zero); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
114 vx7 = vec_mradds (vec_sl (block[7], shift), constants[2], zero); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
115 \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
116 IDCT_HALF \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
117 \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
118 vx0 = vec_mergeh (vy0, vy4); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
119 vx1 = vec_mergel (vy0, vy4); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
120 vx2 = vec_mergeh (vy1, vy5); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
121 vx3 = vec_mergel (vy1, vy5); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
122 vx4 = vec_mergeh (vy2, vy6); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
123 vx5 = vec_mergel (vy2, vy6); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
124 vx6 = vec_mergeh (vy3, vy7); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
125 vx7 = vec_mergel (vy3, vy7); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
126 \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
127 vy0 = vec_mergeh (vx0, vx4); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
128 vy1 = vec_mergel (vx0, vx4); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
129 vy2 = vec_mergeh (vx1, vx5); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
130 vy3 = vec_mergel (vx1, vx5); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
131 vy4 = vec_mergeh (vx2, vx6); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
132 vy5 = vec_mergel (vx2, vx6); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
133 vy6 = vec_mergeh (vx3, vx7); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
134 vy7 = vec_mergel (vx3, vx7); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
135 \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
136 vx0 = vec_adds (vec_mergeh (vy0, vy4), bias); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
137 vx1 = vec_mergel (vy0, vy4); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
138 vx2 = vec_mergeh (vy1, vy5); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
139 vx3 = vec_mergel (vy1, vy5); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
140 vx4 = vec_mergeh (vy2, vy6); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
141 vx5 = vec_mergel (vy2, vy6); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
142 vx6 = vec_mergeh (vy3, vy7); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
143 vx7 = vec_mergel (vy3, vy7); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
144 \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
145 IDCT_HALF \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
146 \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
147 shift = vec_splat_u16 (6); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
148 vx0 = vec_sra (vy0, shift); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
149 vx1 = vec_sra (vy1, shift); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
150 vx2 = vec_sra (vy2, shift); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
151 vx3 = vec_sra (vy3, shift); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
152 vx4 = vec_sra (vy4, shift); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
153 vx5 = vec_sra (vy5, shift); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
154 vx6 = vec_sra (vy6, shift); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
155 vx7 = vec_sra (vy7, shift); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
156 |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
157 |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
158 static const vector_s16_t constants[5] = { |
1277
f3152eb76f1a
altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents:
1064
diff
changeset
|
159 (vector_s16_t) AVV(23170, 13573, 6518, 21895, -23170, -21895, 32, 31), |
f3152eb76f1a
altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents:
1064
diff
changeset
|
160 (vector_s16_t) AVV(16384, 22725, 21407, 19266, 16384, 19266, 21407, 22725), |
f3152eb76f1a
altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents:
1064
diff
changeset
|
161 (vector_s16_t) AVV(22725, 31521, 29692, 26722, 22725, 26722, 29692, 31521), |
f3152eb76f1a
altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents:
1064
diff
changeset
|
162 (vector_s16_t) AVV(21407, 29692, 27969, 25172, 21407, 25172, 27969, 29692), |
f3152eb76f1a
altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents:
1064
diff
changeset
|
163 (vector_s16_t) AVV(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
|
164 }; |
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 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
|
167 { |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1277
diff
changeset
|
168 POWERPC_PERF_DECLARE(altivec_idct_put_num, 1); |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
828
diff
changeset
|
169 #ifdef ALTIVEC_USE_REFERENCE_C_CODE |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1277
diff
changeset
|
170 POWERPC_PERF_START_COUNT(altivec_idct_put_num, 1); |
1064 | 171 void simple_idct_put(uint8_t *dest, int line_size, int16_t *block); |
172 simple_idct_put(dest, stride, (int16_t*)block); | |
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_STOP_COUNT(altivec_idct_put_num, 1); |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
828
diff
changeset
|
174 #else /* ALTIVEC_USE_REFERENCE_C_CODE */ |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
175 vector_u8_t tmp; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
176 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1277
diff
changeset
|
177 POWERPC_PERF_START_COUNT(altivec_idct_put_num, 1); |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
828
diff
changeset
|
178 |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
179 IDCT |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
180 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
181 #define COPY(dest,src) \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
182 tmp = vec_packsu (src, src); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
183 vec_ste ((vector_u32_t)tmp, 0, (unsigned int *)dest); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
184 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
|
185 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
186 COPY (dest, vx0) dest += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
187 COPY (dest, vx1) dest += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
188 COPY (dest, vx2) dest += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
189 COPY (dest, vx3) dest += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
190 COPY (dest, vx4) dest += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
191 COPY (dest, vx5) dest += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
192 COPY (dest, vx6) dest += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
193 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
|
194 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1277
diff
changeset
|
195 POWERPC_PERF_STOP_COUNT(altivec_idct_put_num, 1); |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
828
diff
changeset
|
196 #endif /* ALTIVEC_USE_REFERENCE_C_CODE */ |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
197 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
198 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
199 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
|
200 { |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1277
diff
changeset
|
201 POWERPC_PERF_DECLARE(altivec_idct_add_num, 1); |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
828
diff
changeset
|
202 #ifdef ALTIVEC_USE_REFERENCE_C_CODE |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1277
diff
changeset
|
203 POWERPC_PERF_START_COUNT(altivec_idct_add_num, 1); |
1064 | 204 void simple_idct_add(uint8_t *dest, int line_size, int16_t *block); |
205 simple_idct_add(dest, stride, (int16_t*)block); | |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1277
diff
changeset
|
206 POWERPC_PERF_STOP_COUNT(altivec_idct_add_num, 1); |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
828
diff
changeset
|
207 #else /* ALTIVEC_USE_REFERENCE_C_CODE */ |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
208 vector_u8_t tmp; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
209 vector_s16_t tmp2, tmp3; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
210 vector_u8_t perm0; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
211 vector_u8_t perm1; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
212 vector_u8_t p0, p1, p; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
213 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1277
diff
changeset
|
214 POWERPC_PERF_START_COUNT(altivec_idct_add_num, 1); |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
828
diff
changeset
|
215 |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
216 IDCT |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
217 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
218 p0 = vec_lvsl (0, dest); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
219 p1 = vec_lvsl (stride, dest); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
220 p = vec_splat_u8 (-1); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
221 perm0 = vec_mergeh (p, p0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
222 perm1 = vec_mergeh (p, p1); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
223 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
224 #define ADD(dest,src,perm) \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
225 /* *(uint64_t *)&tmp = *(uint64_t *)dest; */ \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
226 tmp = vec_ld (0, dest); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
227 tmp2 = (vector_s16_t)vec_perm (tmp, (vector_u8_t)zero, perm); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
228 tmp3 = vec_adds (tmp2, src); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
229 tmp = vec_packsu (tmp3, tmp3); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
230 vec_ste ((vector_u32_t)tmp, 0, (unsigned int *)dest); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
231 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
|
232 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
233 ADD (dest, vx0, perm0) dest += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
234 ADD (dest, vx1, perm1) dest += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
235 ADD (dest, vx2, perm0) dest += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
236 ADD (dest, vx3, perm1) dest += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
237 ADD (dest, vx4, perm0) dest += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
238 ADD (dest, vx5, perm1) dest += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
239 ADD (dest, vx6, perm0) dest += stride; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
240 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
|
241 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1277
diff
changeset
|
242 POWERPC_PERF_STOP_COUNT(altivec_idct_add_num, 1); |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
828
diff
changeset
|
243 #endif /* ALTIVEC_USE_REFERENCE_C_CODE */ |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
244 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
245 |