4 * Copyright (c) 2009 Loren Merritt
6 * This file is part of FFmpeg.
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.
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.
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
24 #include "libavutil/cpu.h"
25 #include "libavutil/ppc/cpu.h"
26 #include "libavutil/ppc/util_altivec.h"
27 #include "libavcodec/fft.h"
30 * Do a complex FFT with the parameters defined in ff_fft_init().
31 * The input data must be permuted before with s->revtab table.
32 * No 1.0 / sqrt(n) normalization is done.
34 * This code assumes that the 'z' pointer is 16 bytes-aligned.
35 * It also assumes all FFTComplex are 8 bytes-aligned pairs of floats.
41 void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z);
42 void ff_fft_calc_interleave_altivec(FFTContext *s, FFTComplex *z);
45 #if HAVE_GNU_AS && HAVE_ALTIVEC && (HAVE_BIGENDIAN || HAVE_VSX)
46 static void imdct_half_altivec(FFTContext *s, FFTSample *output, const FFTSample *input)
49 int n = 1 << s->mdct_bits;
53 const uint16_t *revtabj = s->revtab;
54 const uint16_t *revtabk = s->revtab+n4;
55 const vec_f *tcos = (const vec_f*)(s->tcos+n8);
56 const vec_f *tsin = (const vec_f*)(s->tsin+n8);
57 const vec_f *pin = (const vec_f*)(input+n4);
58 vec_f *pout = (vec_f*)(output+n4);
63 vec_f cos,sin,cos0,sin0,cos1,sin1,re,im,r0,i0,r1,i1,a,b,c,d;
64 #define CMULA(p,o0,o1,o2,o3)\
65 a = pin[ k*2+p]; /* { z[k].re, z[k].im, z[k+1].re, z[k+1].im } */\
66 b = pin[-k*2-p-1]; /* { z[-k-2].re, z[-k-2].im, z[-k-1].re, z[-k-1].im } */\
67 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 } */\
68 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 } */\
69 cos = vec_perm(cos0, cos1, vcprm(o0,o1,s##o2,s##o3)); /* { cos[k], cos[k+1], cos[-k-2], cos[-k-1] } */\
70 sin = vec_perm(sin0, sin1, vcprm(o0,o1,s##o2,s##o3));\
71 r##p = im*cos - re*sin;\
72 i##p = re*cos + im*sin;
73 #define STORE2(v,dst)\
75 vec_ste(v, 0, output+j*2);\
76 vec_ste(v, 4, output+j*2);
78 a = vec_perm(r##p, i##p, vcprm(0,s0,0,s0));\
79 b = vec_perm(r##p, i##p, vcprm(1,s1,1,s1));\
80 c = vec_perm(r##p, i##p, vcprm(2,s2,2,s2));\
81 d = vec_perm(r##p, i##p, vcprm(3,s3,3,s3));\
82 STORE2(a, revtabk[ p*2-4]);\
83 STORE2(b, revtabk[ p*2-3]);\
84 STORE2(c, revtabj[-p*2+2]);\
85 STORE2(d, revtabj[-p*2+3]);
101 ff_fft_calc_vsx(s, (FFTComplex*)output);
103 ff_fft_calc_altivec(s, (FFTComplex*)output);
106 /* post rotation + reordering */
110 vec_f cos,sin,re,im,a,b,c,d;
111 #define CMULB(d0,d1,o)\
116 d0 = im*sin - re*cos;\
117 d1 = re*sin + im*cos;
121 pout[2*j] = vec_perm(a, d, vcprm(0,s3,1,s2));
122 pout[2*j+1] = vec_perm(a, d, vcprm(2,s1,3,s0));
123 pout[2*k] = vec_perm(c, b, vcprm(0,s3,1,s2));
124 pout[2*k+1] = vec_perm(c, b, vcprm(2,s1,3,s0));
130 static void imdct_calc_altivec(FFTContext *s, FFTSample *output, const FFTSample *input)
133 int n = 1 << s->mdct_bits;
136 vec_u32 sign = {1U<<31,1U<<31,1U<<31,1U<<31};
137 vec_u32 *p0 = (vec_u32*)(output+n4);
138 vec_u32 *p1 = (vec_u32*)(output+n4*3);
140 imdct_half_altivec(s, output + n4, input);
142 for (k = 0; k < n16; k++) {
143 vec_u32 a = p0[k] ^ sign;
144 vec_u32 b = p1[-k-1];
145 p0[-k-1] = vec_perm(a, a, vcprm(3,2,1,0));
146 p1[k] = vec_perm(b, b, vcprm(3,2,1,0));
149 #endif /* HAVE_GNU_AS && HAVE_ALTIVEC && (HAVE_BIGENDIAN || HAVE_VSX) */
151 av_cold void ff_fft_init_ppc(FFTContext *s)
153 #if HAVE_GNU_AS && HAVE_ALTIVEC && (HAVE_BIGENDIAN || HAVE_VSX)
154 if (!PPC_ALTIVEC(av_get_cpu_flags()))
158 s->fft_calc = ff_fft_calc_interleave_vsx;
160 s->fft_calc = ff_fft_calc_interleave_altivec;
162 if (s->mdct_bits >= 5) {
163 s->imdct_calc = imdct_calc_altivec;
164 s->imdct_half = imdct_half_altivec;
166 #endif /* HAVE_GNU_AS && HAVE_ALTIVEC && HAVE_BIGENDIAN */