1/*
2 * FFT/IFFT transforms
3 * AltiVec-enabled
4 * Copyright (c) 2009 Loren Merritt
5 *
6 * This file is part of Libav.
7 *
8 * Libav 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 * Libav 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 Libav; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 */
22#include "libavcodec/fft.h"
23#include "util_altivec.h"
24#include "types_altivec.h"
25
26/**
27 * Do a complex FFT with the parameters defined in ff_fft_init(). The
28 * input data must be permuted before with s->revtab table. No
29 * 1.0/sqrt(n) normalization is done.
30 * AltiVec-enabled
31 * This code assumes that the 'z' pointer is 16 bytes-aligned
32 * It also assumes all FFTComplex are 8 bytes-aligned pair of float
33 */
34
35void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z);
36void ff_fft_calc_interleave_altivec(FFTContext *s, FFTComplex *z);
37
38#if HAVE_GNU_AS
39static void ff_imdct_half_altivec(FFTContext *s, FFTSample *output, const FFTSample *input)
40{
41    int j, k;
42    int n = 1 << s->mdct_bits;
43    int n4 = n >> 2;
44    int n8 = n >> 3;
45    int n32 = n >> 5;
46    const uint16_t *revtabj = s->revtab;
47    const uint16_t *revtabk = s->revtab+n4;
48    const vec_f *tcos = (const vec_f*)(s->tcos+n8);
49    const vec_f *tsin = (const vec_f*)(s->tsin+n8);
50    const vec_f *pin = (const vec_f*)(input+n4);
51    vec_f *pout = (vec_f*)(output+n4);
52
53    /* pre rotation */
54    k = n32-1;
55    do {
56        vec_f cos,sin,cos0,sin0,cos1,sin1,re,im,r0,i0,r1,i1,a,b,c,d;
57#define CMULA(p,o0,o1,o2,o3)\
58        a = pin[ k*2+p];                       /* { z[k].re,    z[k].im,    z[k+1].re,  z[k+1].im  } */\
59        b = pin[-k*2-p-1];                     /* { z[-k-2].re, z[-k-2].im, z[-k-1].re, z[-k-1].im } */\
60        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 } */\
61        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    } */\
62        cos = vec_perm(cos0, cos1, vcprm(o0,o1,s##o2,s##o3)); /* { cos[k], cos[k+1], cos[-k-2], cos[-k-1] } */\
63        sin = vec_perm(sin0, sin1, vcprm(o0,o1,s##o2,s##o3));\
64        r##p = im*cos - re*sin;\
65        i##p = re*cos + im*sin;
66#define STORE2(v,dst)\
67        j = dst;\
68        vec_ste(v, 0, output+j*2);\
69        vec_ste(v, 4, output+j*2);
70#define STORE8(p)\
71        a = vec_perm(r##p, i##p, vcprm(0,s0,0,s0));\
72        b = vec_perm(r##p, i##p, vcprm(1,s1,1,s1));\
73        c = vec_perm(r##p, i##p, vcprm(2,s2,2,s2));\
74        d = vec_perm(r##p, i##p, vcprm(3,s3,3,s3));\
75        STORE2(a, revtabk[ p*2-4]);\
76        STORE2(b, revtabk[ p*2-3]);\
77        STORE2(c, revtabj[-p*2+2]);\
78        STORE2(d, revtabj[-p*2+3]);
79
80        cos0 = tcos[k];
81        sin0 = tsin[k];
82        cos1 = tcos[-k-1];
83        sin1 = tsin[-k-1];
84        CMULA(0, 0,1,2,3);
85        CMULA(1, 2,3,0,1);
86        STORE8(0);
87        STORE8(1);
88        revtabj += 4;
89        revtabk -= 4;
90        k--;
91    } while(k >= 0);
92
93    ff_fft_calc_altivec(s, (FFTComplex*)output);
94
95    /* post rotation + reordering */
96    j = -n32;
97    k = n32-1;
98    do {
99        vec_f cos,sin,re,im,a,b,c,d;
100#define CMULB(d0,d1,o)\
101        re = pout[o*2];\
102        im = pout[o*2+1];\
103        cos = tcos[o];\
104        sin = tsin[o];\
105        d0 = im*sin - re*cos;\
106        d1 = re*sin + im*cos;
107
108        CMULB(a,b,j);
109        CMULB(c,d,k);
110        pout[2*j]   = vec_perm(a, d, vcprm(0,s3,1,s2));
111        pout[2*j+1] = vec_perm(a, d, vcprm(2,s1,3,s0));
112        pout[2*k]   = vec_perm(c, b, vcprm(0,s3,1,s2));
113        pout[2*k+1] = vec_perm(c, b, vcprm(2,s1,3,s0));
114        j++;
115        k--;
116    } while(k >= 0);
117}
118
119static void ff_imdct_calc_altivec(FFTContext *s, FFTSample *output, const FFTSample *input)
120{
121    int k;
122    int n = 1 << s->mdct_bits;
123    int n4 = n >> 2;
124    int n16 = n >> 4;
125    vec_u32 sign = {1U<<31,1U<<31,1U<<31,1U<<31};
126    vec_u32 *p0 = (vec_u32*)(output+n4);
127    vec_u32 *p1 = (vec_u32*)(output+n4*3);
128
129    ff_imdct_half_altivec(s, output+n4, input);
130
131    for (k = 0; k < n16; k++) {
132        vec_u32 a = p0[k] ^ sign;
133        vec_u32 b = p1[-k-1];
134        p0[-k-1] = vec_perm(a, a, vcprm(3,2,1,0));
135        p1[k]    = vec_perm(b, b, vcprm(3,2,1,0));
136    }
137}
138#endif /* HAVE_GNU_AS */
139
140av_cold void ff_fft_init_altivec(FFTContext *s)
141{
142#if HAVE_GNU_AS
143    s->fft_calc   = ff_fft_calc_interleave_altivec;
144    if (s->mdct_bits >= 5) {
145        s->imdct_calc = ff_imdct_calc_altivec;
146        s->imdct_half = ff_imdct_half_altivec;
147    }
148#endif
149}
150