Mercurial > libavcodec.hg
annotate ppc/util_altivec.h @ 12046:ae57be2ef58c libavcodec
PPC: Altivec split-radix FFT
1.8x faster than altivec radix-2 on a G4
8% faster vorbis decoding
Patch (mostly) by Loren Merritt
author | mru |
---|---|
date | Thu, 01 Jul 2010 23:21:39 +0000 |
parents | 8499462c732f |
children |
rev | line source |
---|---|
5750
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
1 /* |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
2 * This file is part of FFmpeg. |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
3 * |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
4 * FFmpeg is free software; you can redistribute it and/or |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
5 * modify it under the terms of the GNU Lesser General Public |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
6 * License as published by the Free Software Foundation; either |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
7 * version 2.1 of the License, or (at your option) any later version. |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
8 * |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
9 * FFmpeg is distributed in the hope that it will be useful, |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
12 * Lesser General Public License for more details. |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
13 * |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
14 * You should have received a copy of the GNU Lesser General Public |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
15 * License along with FFmpeg; if not, write to the Free Software |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
16 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
17 */ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
18 |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
19 /** |
11644
7dd2a45249a9
Remove explicit filename from Doxygen @file commands.
diego
parents:
8718
diff
changeset
|
20 * @file |
5750
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
21 * Contains misc utility macros and inline functions |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
22 */ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
23 |
7760 | 24 #ifndef AVCODEC_PPC_UTIL_ALTIVEC_H |
25 #define AVCODEC_PPC_UTIL_ALTIVEC_H | |
5750
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
26 |
6078 | 27 #include <stdint.h> |
28 | |
29 #include "config.h" | |
30 | |
8590 | 31 #if HAVE_ALTIVEC_H |
6078 | 32 #include <altivec.h> |
33 #endif | |
34 | |
12010
8499462c732f
Altivec: Add helper function to load from a constant misalignment
conrad
parents:
11644
diff
changeset
|
35 #include "types_altivec.h" |
8499462c732f
Altivec: Add helper function to load from a constant misalignment
conrad
parents:
11644
diff
changeset
|
36 |
5750
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
37 // used to build registers permutation vectors (vcprm) |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
38 // the 's' are for words in the _s_econd vector |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
39 #define WORD_0 0x00,0x01,0x02,0x03 |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
40 #define WORD_1 0x04,0x05,0x06,0x07 |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
41 #define WORD_2 0x08,0x09,0x0a,0x0b |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
42 #define WORD_3 0x0c,0x0d,0x0e,0x0f |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
43 #define WORD_s0 0x10,0x11,0x12,0x13 |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
44 #define WORD_s1 0x14,0x15,0x16,0x17 |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
45 #define WORD_s2 0x18,0x19,0x1a,0x1b |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
46 #define WORD_s3 0x1c,0x1d,0x1e,0x1f |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
47 |
7373
266d4949aa15
Remove AltiVec vector declaration compiler compatibility macros.
diego
parents:
6368
diff
changeset
|
48 #define vcprm(a,b,c,d) (const vector unsigned char){WORD_ ## a, WORD_ ## b, WORD_ ## c, WORD_ ## d} |
266d4949aa15
Remove AltiVec vector declaration compiler compatibility macros.
diego
parents:
6368
diff
changeset
|
49 #define vcii(a,b,c,d) (const vector float){FLOAT_ ## a, FLOAT_ ## b, FLOAT_ ## c, FLOAT_ ## d} |
5750
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
50 |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
51 // vcprmle is used to keep the same index as in the SSE version. |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
52 // it's the same as vcprm, with the index inversed |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
53 // ('le' is Little Endian) |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
54 #define vcprmle(a,b,c,d) vcprm(d,c,b,a) |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
55 |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
56 // used to build inverse/identity vectors (vcii) |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
57 // n is _n_egative, p is _p_ositive |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
58 #define FLOAT_n -1. |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
59 #define FLOAT_p 1. |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
60 |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
61 |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
62 // Transpose 8x8 matrix of 16-bit elements (in-place) |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
63 #define TRANSPOSE8(a,b,c,d,e,f,g,h) \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
64 do { \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
65 vector signed short A1, B1, C1, D1, E1, F1, G1, H1; \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
66 vector signed short A2, B2, C2, D2, E2, F2, G2, H2; \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
67 \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
68 A1 = vec_mergeh (a, e); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
69 B1 = vec_mergel (a, e); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
70 C1 = vec_mergeh (b, f); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
71 D1 = vec_mergel (b, f); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
72 E1 = vec_mergeh (c, g); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
73 F1 = vec_mergel (c, g); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
74 G1 = vec_mergeh (d, h); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
75 H1 = vec_mergel (d, h); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
76 \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
77 A2 = vec_mergeh (A1, E1); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
78 B2 = vec_mergel (A1, E1); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
79 C2 = vec_mergeh (B1, F1); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
80 D2 = vec_mergel (B1, F1); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
81 E2 = vec_mergeh (C1, G1); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
82 F2 = vec_mergel (C1, G1); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
83 G2 = vec_mergeh (D1, H1); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
84 H2 = vec_mergel (D1, H1); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
85 \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
86 a = vec_mergeh (A2, E2); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
87 b = vec_mergel (A2, E2); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
88 c = vec_mergeh (B2, F2); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
89 d = vec_mergel (B2, F2); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
90 e = vec_mergeh (C2, G2); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
91 f = vec_mergel (C2, G2); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
92 g = vec_mergeh (D2, H2); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
93 h = vec_mergel (D2, H2); \ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
94 } while (0) |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
95 |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
96 |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
97 /** \brief loads unaligned vector \a *src with offset \a offset |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
98 and returns it */ |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
99 static inline vector unsigned char unaligned_load(int offset, uint8_t *src) |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
100 { |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
101 register vector unsigned char first = vec_ld(offset, src); |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
102 register vector unsigned char second = vec_ld(offset+15, src); |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
103 register vector unsigned char mask = vec_lvsl(offset, src); |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
104 return vec_perm(first, second, mask); |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
105 } |
09f99af1db40
Sanitize altivec code so it can be built with runtime check properly
lu_zero
parents:
diff
changeset
|
106 |
12010
8499462c732f
Altivec: Add helper function to load from a constant misalignment
conrad
parents:
11644
diff
changeset
|
107 /** |
8499462c732f
Altivec: Add helper function to load from a constant misalignment
conrad
parents:
11644
diff
changeset
|
108 * loads vector known misalignment |
8499462c732f
Altivec: Add helper function to load from a constant misalignment
conrad
parents:
11644
diff
changeset
|
109 * @param perm_vec the align permute vector to combine the two loads from lvsl |
8499462c732f
Altivec: Add helper function to load from a constant misalignment
conrad
parents:
11644
diff
changeset
|
110 */ |
8499462c732f
Altivec: Add helper function to load from a constant misalignment
conrad
parents:
11644
diff
changeset
|
111 static inline vec_u8 load_with_perm_vec(int offset, uint8_t *src, vec_u8 perm_vec) |
8499462c732f
Altivec: Add helper function to load from a constant misalignment
conrad
parents:
11644
diff
changeset
|
112 { |
8499462c732f
Altivec: Add helper function to load from a constant misalignment
conrad
parents:
11644
diff
changeset
|
113 vec_u8 a = vec_ld(offset, src); |
8499462c732f
Altivec: Add helper function to load from a constant misalignment
conrad
parents:
11644
diff
changeset
|
114 vec_u8 b = vec_ld(offset+15, src); |
8499462c732f
Altivec: Add helper function to load from a constant misalignment
conrad
parents:
11644
diff
changeset
|
115 return vec_perm(a, b, perm_vec); |
8499462c732f
Altivec: Add helper function to load from a constant misalignment
conrad
parents:
11644
diff
changeset
|
116 } |
8499462c732f
Altivec: Add helper function to load from a constant misalignment
conrad
parents:
11644
diff
changeset
|
117 |
7760 | 118 #endif /* AVCODEC_PPC_UTIL_ALTIVEC_H */ |