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"
24 #include "libavutil/cpu.h"
27 #include "libavcodec/fft.h"
28 
29 /**
30  * Do a complex FFT with the parameters defined in ff_fft_init(). The
31  * input data must be permuted before with s->revtab table. No
32  * 1.0/sqrt(n) normalization is done.
33  * AltiVec-enabled
34  * This code assumes that the 'z' pointer is 16 bytes-aligned
35  * It also assumes all FFTComplex are 8 bytes-aligned pair of float
36  */
37 
40 
41 #if HAVE_GNU_AS && HAVE_ALTIVEC
42 static void imdct_half_altivec(FFTContext *s, FFTSample *output, const FFTSample *input)
43 {
44  int j, k;
45  int n = 1 << s->mdct_bits;
46  int n4 = n >> 2;
47  int n8 = n >> 3;
48  int n32 = n >> 5;
49  const uint16_t *revtabj = s->revtab;
50  const uint16_t *revtabk = s->revtab+n4;
51  const vec_f *tcos = (const vec_f*)(s->tcos+n8);
52  const vec_f *tsin = (const vec_f*)(s->tsin+n8);
53  const vec_f *pin = (const vec_f*)(input+n4);
54  vec_f *pout = (vec_f*)(output+n4);
55 
56  /* pre rotation */
57  k = n32-1;
58  do {
59  vec_f cos,sin,cos0,sin0,cos1,sin1,re,im,r0,i0,r1,i1,a,b,c,d;
60 #define CMULA(p,o0,o1,o2,o3)\
61  a = pin[ k*2+p]; /* { z[k].re, z[k].im, z[k+1].re, z[k+1].im } */\
62  b = pin[-k*2-p-1]; /* { z[-k-2].re, z[-k-2].im, z[-k-1].re, z[-k-1].im } */\
63  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 } */\
64  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 } */\
65  cos = vec_perm(cos0, cos1, vcprm(o0,o1,s##o2,s##o3)); /* { cos[k], cos[k+1], cos[-k-2], cos[-k-1] } */\
66  sin = vec_perm(sin0, sin1, vcprm(o0,o1,s##o2,s##o3));\
67  r##p = im*cos - re*sin;\
68  i##p = re*cos + im*sin;
69 #define STORE2(v,dst)\
70  j = dst;\
71  vec_ste(v, 0, output+j*2);\
72  vec_ste(v, 4, output+j*2);
73 #define STORE8(p)\
74  a = vec_perm(r##p, i##p, vcprm(0,s0,0,s0));\
75  b = vec_perm(r##p, i##p, vcprm(1,s1,1,s1));\
76  c = vec_perm(r##p, i##p, vcprm(2,s2,2,s2));\
77  d = vec_perm(r##p, i##p, vcprm(3,s3,3,s3));\
78  STORE2(a, revtabk[ p*2-4]);\
79  STORE2(b, revtabk[ p*2-3]);\
80  STORE2(c, revtabj[-p*2+2]);\
81  STORE2(d, revtabj[-p*2+3]);
82 
83  cos0 = tcos[k];
84  sin0 = tsin[k];
85  cos1 = tcos[-k-1];
86  sin1 = tsin[-k-1];
87  CMULA(0, 0,1,2,3);
88  CMULA(1, 2,3,0,1);
89  STORE8(0);
90  STORE8(1);
91  revtabj += 4;
92  revtabk -= 4;
93  k--;
94  } while(k >= 0);
95 
96  ff_fft_calc_altivec(s, (FFTComplex*)output);
97 
98  /* post rotation + reordering */
99  j = -n32;
100  k = n32-1;
101  do {
102  vec_f cos,sin,re,im,a,b,c,d;
103 #define CMULB(d0,d1,o)\
104  re = pout[o*2];\
105  im = pout[o*2+1];\
106  cos = tcos[o];\
107  sin = tsin[o];\
108  d0 = im*sin - re*cos;\
109  d1 = re*sin + im*cos;
110 
111  CMULB(a,b,j);
112  CMULB(c,d,k);
113  pout[2*j] = vec_perm(a, d, vcprm(0,s3,1,s2));
114  pout[2*j+1] = vec_perm(a, d, vcprm(2,s1,3,s0));
115  pout[2*k] = vec_perm(c, b, vcprm(0,s3,1,s2));
116  pout[2*k+1] = vec_perm(c, b, vcprm(2,s1,3,s0));
117  j++;
118  k--;
119  } while(k >= 0);
120 }
121 
122 static void imdct_calc_altivec(FFTContext *s, FFTSample *output, const FFTSample *input)
123 {
124  int k;
125  int n = 1 << s->mdct_bits;
126  int n4 = n >> 2;
127  int n16 = n >> 4;
128  vec_u32 sign = {1U<<31,1U<<31,1U<<31,1U<<31};
129  vec_u32 *p0 = (vec_u32*)(output+n4);
130  vec_u32 *p1 = (vec_u32*)(output+n4*3);
131 
132  imdct_half_altivec(s, output + n4, input);
133 
134  for (k = 0; k < n16; k++) {
135  vec_u32 a = p0[k] ^ sign;
136  vec_u32 b = p1[-k-1];
137  p0[-k-1] = vec_perm(a, a, vcprm(3,2,1,0));
138  p1[k] = vec_perm(b, b, vcprm(3,2,1,0));
139  }
140 }
141 #endif /* HAVE_GNU_AS && HAVE_ALTIVEC */
142 
144 {
145 #if HAVE_GNU_AS && HAVE_ALTIVEC
147  return;
148 
150  if (s->mdct_bits >= 5) {
151  s->imdct_calc = imdct_calc_altivec;
152  s->imdct_half = imdct_half_altivec;
153  }
154 #endif /* HAVE_GNU_AS && HAVE_ALTIVEC */
155 }