FFmpeg
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
fft_altivec.c
Go to the documentation of this file.
1 /*
2  * FFT/IFFT transforms
3  * AltiVec-enabled
4  * Copyright (c) 2009 Loren Merritt
5  *
6  * This file is part of FFmpeg.
7  *
8  * FFmpeg is free software; you can redistribute it and/or
9  * modify it under the terms of the GNU Lesser General Public
10  * License as published by the Free Software Foundation; either
11  * version 2.1 of the License, or (at your option) any later version.
12  *
13  * FFmpeg is distributed in the hope that it will be useful,
14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16  * Lesser General Public License for more details.
17  *
18  * You should have received a copy of the GNU Lesser General Public
19  * License along with FFmpeg; if not, write to the Free Software
20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21  */
22 
23 #include "config.h"
26 #include "libavcodec/fft.h"
27 
28 /**
29  * Do a complex FFT with the parameters defined in ff_fft_init(). The
30  * input data must be permuted before with s->revtab table. No
31  * 1.0/sqrt(n) normalization is done.
32  * AltiVec-enabled
33  * This code assumes that the 'z' pointer is 16 bytes-aligned
34  * It also assumes all FFTComplex are 8 bytes-aligned pair of float
35  */
36 
39 
40 #if HAVE_GNU_AS && HAVE_ALTIVEC
41 static void imdct_half_altivec(FFTContext *s, FFTSample *output, const FFTSample *input)
42 {
43  int j, k;
44  int n = 1 << s->mdct_bits;
45  int n4 = n >> 2;
46  int n8 = n >> 3;
47  int n32 = n >> 5;
48  const uint16_t *revtabj = s->revtab;
49  const uint16_t *revtabk = s->revtab+n4;
50  const vec_f *tcos = (const vec_f*)(s->tcos+n8);
51  const vec_f *tsin = (const vec_f*)(s->tsin+n8);
52  const vec_f *pin = (const vec_f*)(input+n4);
53  vec_f *pout = (vec_f*)(output+n4);
54 
55  /* pre rotation */
56  k = n32-1;
57  do {
58  vec_f cos,sin,cos0,sin0,cos1,sin1,re,im,r0,i0,r1,i1,a,b,c,d;
59 #define CMULA(p,o0,o1,o2,o3)\
60  a = pin[ k*2+p]; /* { z[k].re, z[k].im, z[k+1].re, z[k+1].im } */\
61  b = pin[-k*2-p-1]; /* { z[-k-2].re, z[-k-2].im, z[-k-1].re, z[-k-1].im } */\
62  re = vec_perm(a, b, vcprm(0,2,s0,s2)); /* { z[k].re, z[k+1].re, z[-k-2].re, z[-k-1].re } */\
63  im = vec_perm(a, b, vcprm(s3,s1,3,1)); /* { z[-k-1].im, z[-k-2].im, z[k+1].im, z[k].im } */\
64  cos = vec_perm(cos0, cos1, vcprm(o0,o1,s##o2,s##o3)); /* { cos[k], cos[k+1], cos[-k-2], cos[-k-1] } */\
65  sin = vec_perm(sin0, sin1, vcprm(o0,o1,s##o2,s##o3));\
66  r##p = im*cos - re*sin;\
67  i##p = re*cos + im*sin;
68 #define STORE2(v,dst)\
69  j = dst;\
70  vec_ste(v, 0, output+j*2);\
71  vec_ste(v, 4, output+j*2);
72 #define STORE8(p)\
73  a = vec_perm(r##p, i##p, vcprm(0,s0,0,s0));\
74  b = vec_perm(r##p, i##p, vcprm(1,s1,1,s1));\
75  c = vec_perm(r##p, i##p, vcprm(2,s2,2,s2));\
76  d = vec_perm(r##p, i##p, vcprm(3,s3,3,s3));\
77  STORE2(a, revtabk[ p*2-4]);\
78  STORE2(b, revtabk[ p*2-3]);\
79  STORE2(c, revtabj[-p*2+2]);\
80  STORE2(d, revtabj[-p*2+3]);
81 
82  cos0 = tcos[k];
83  sin0 = tsin[k];
84  cos1 = tcos[-k-1];
85  sin1 = tsin[-k-1];
86  CMULA(0, 0,1,2,3);
87  CMULA(1, 2,3,0,1);
88  STORE8(0);
89  STORE8(1);
90  revtabj += 4;
91  revtabk -= 4;
92  k--;
93  } while(k >= 0);
94 
95  ff_fft_calc_altivec(s, (FFTComplex*)output);
96 
97  /* post rotation + reordering */
98  j = -n32;
99  k = n32-1;
100  do {
101  vec_f cos,sin,re,im,a,b,c,d;
102 #define CMULB(d0,d1,o)\
103  re = pout[o*2];\
104  im = pout[o*2+1];\
105  cos = tcos[o];\
106  sin = tsin[o];\
107  d0 = im*sin - re*cos;\
108  d1 = re*sin + im*cos;
109 
110  CMULB(a,b,j);
111  CMULB(c,d,k);
112  pout[2*j] = vec_perm(a, d, vcprm(0,s3,1,s2));
113  pout[2*j+1] = vec_perm(a, d, vcprm(2,s1,3,s0));
114  pout[2*k] = vec_perm(c, b, vcprm(0,s3,1,s2));
115  pout[2*k+1] = vec_perm(c, b, vcprm(2,s1,3,s0));
116  j++;
117  k--;
118  } while(k >= 0);
119 }
120 
121 static void imdct_calc_altivec(FFTContext *s, FFTSample *output, const FFTSample *input)
122 {
123  int k;
124  int n = 1 << s->mdct_bits;
125  int n4 = n >> 2;
126  int n16 = n >> 4;
127  vec_u32 sign = {1U<<31,1U<<31,1U<<31,1U<<31};
128  vec_u32 *p0 = (vec_u32*)(output+n4);
129  vec_u32 *p1 = (vec_u32*)(output+n4*3);
130 
131  imdct_half_altivec(s, output + n4, input);
132 
133  for (k = 0; k < n16; k++) {
134  vec_u32 a = p0[k] ^ sign;
135  vec_u32 b = p1[-k-1];
136  p0[-k-1] = vec_perm(a, a, vcprm(3,2,1,0));
137  p1[k] = vec_perm(b, b, vcprm(3,2,1,0));
138  }
139 }
140 #endif /* HAVE_GNU_AS && HAVE_ALTIVEC */
141 
143 {
144 #if HAVE_GNU_AS && HAVE_ALTIVEC
146  if (s->mdct_bits >= 5) {
147  s->imdct_calc = imdct_calc_altivec;
148  s->imdct_half = imdct_half_altivec;
149  }
150 #endif /* HAVE_GNU_AS && HAVE_ALTIVEC */
151 }