Mercurial > libavcodec.hg
comparison ppc/fft_altivec.c @ 981:8bec850dc9c7 libavcodec
altivec patches by Romain Dolbeau
author | bellard |
---|---|
date | Wed, 08 Jan 2003 18:47:49 +0000 |
parents | e05d525505c5 |
children | edc10966b081 |
comparison
equal
deleted
inserted
replaced
980:0c32f81b42b2 | 981:8bec850dc9c7 |
---|---|
1 /* | 1 /* |
2 * FFT/IFFT transforms | 2 * FFT/IFFT transforms |
3 * AltiVec-enabled | 3 * AltiVec-enabled |
4 * Copyright (c) 2002 Romain Dolbeau <romain@dolbeau.org> | 4 * Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org> |
5 * Based on code Copyright (c) 2002 Fabrice Bellard. | 5 * Based on code Copyright (c) 2002 Fabrice Bellard. |
6 * | 6 * |
7 * This library is free software; you can redistribute it and/or | 7 * This library is free software; you can redistribute it and/or |
8 * modify it under the terms of the GNU Lesser General Public | 8 * modify it under the terms of the GNU Lesser General Public |
9 * License as published by the Free Software Foundation; either | 9 * License as published by the Free Software Foundation; either |
20 */ | 20 */ |
21 #include "../dsputil.h" | 21 #include "../dsputil.h" |
22 | 22 |
23 #include "dsputil_altivec.h" | 23 #include "dsputil_altivec.h" |
24 | 24 |
25 // used to build registers permutation vectors (vcprm) | |
26 // the 's' are for words in the _s_econd vector | |
27 #define WORD_0 0x00,0x01,0x02,0x03 | |
28 #define WORD_1 0x04,0x05,0x06,0x07 | |
29 #define WORD_2 0x08,0x09,0x0a,0x0b | |
30 #define WORD_3 0x0c,0x0d,0x0e,0x0f | |
31 #define WORD_s0 0x10,0x11,0x12,0x13 | |
32 #define WORD_s1 0x14,0x15,0x16,0x17 | |
33 #define WORD_s2 0x18,0x19,0x1a,0x1b | |
34 #define WORD_s3 0x1c,0x1d,0x1e,0x1f | |
35 | |
36 #define vcprm(a,b,c,d) (const vector unsigned char)(WORD_ ## a, WORD_ ## b, WORD_ ## c, WORD_ ## d) | |
37 | |
38 // vcprmle is used to keep the same index as in the SSE version. | |
39 // it's the same as vcprm, with the index inversed | |
40 // ('le' is Little Endian) | |
41 #define vcprmle(a,b,c,d) vcprm(d,c,b,a) | |
42 | |
43 // used to build inverse/identity vectors (vcii) | |
44 // n is _n_egative, p is _p_ositive | |
45 #define FLOAT_n -1. | |
46 #define FLOAT_p 1. | |
47 | |
48 #define vcii(a,b,c,d) (const vector float)(FLOAT_ ## a, FLOAT_ ## b, FLOAT_ ## c, FLOAT_ ## d) | |
49 | |
50 /** | 25 /** |
51 * Do a complex FFT with the parameters defined in fft_init(). The | 26 * Do a complex FFT with the parameters defined in fft_init(). The |
52 * input data must be permuted before with s->revtab table. No | 27 * input data must be permuted before with s->revtab table. No |
53 * 1.0/sqrt(n) normalization is done. | 28 * 1.0/sqrt(n) normalization is done. |
54 * AltiVec-enabled | 29 * AltiVec-enabled |
55 * This code assumes that the 'z' pointer is 16 bytes-aligned | 30 * This code assumes that the 'z' pointer is 16 bytes-aligned |
56 * It also assumes all FFTComplex are 8 bytes-aligned pair of float | 31 * It also assumes all FFTComplex are 8 bytes-aligned pair of float |
57 * The code is exactly the same as the SSE version, except | 32 * The code is exactly the same as the SSE version, except |
58 * that successive MUL + ADD/SUB have been fusionned into | 33 * that successive MUL + ADD/SUB have been merged into |
59 * fused multiply-add ('vec_madd' in altivec) | 34 * fused multiply-add ('vec_madd' in altivec) |
60 * | |
61 * To test this code you can use fft-test in libavcodec ; use | |
62 * the following line in libavcodec to compile (MacOS X): | |
63 * ##### | |
64 * gcc -I. -Ippc -no-cpp-precomp -pipe -O3 -fomit-frame-pointer -mdynamic-no-pic -Wall | |
65 * -faltivec -DARCH_POWERPC -DHAVE_ALTIVEC -DCONFIG_DARWIN fft-test.c fft.c | |
66 * ppc/fft_altivec.c ppc/dsputil_altivec.c mdct.c -DHAVE_LRINTF -o fft-test | |
67 * ##### | |
68 */ | 35 */ |
69 void fft_calc_altivec(FFTContext *s, FFTComplex *z) | 36 void fft_calc_altivec(FFTContext *s, FFTComplex *z) |
70 { | 37 { |
71 register const vector float vczero = (vector float)( 0., 0., 0., 0.); | 38 register const vector float vczero = (vector float)( 0., 0., 0., 0.); |
72 | 39 |