libavcodec/ppc/fft_init.c

   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
  23 #include "config.h"
  24 #include "libavutil/cpu.h"
  25 #include "libavutil/ppc/cpu.h"
  26 #include "libavutil/ppc/types_altivec.h"
  27 #include "libavutil/ppc/util_altivec.h"
  28 #include "libavcodec/fft.h"
  29
  30 /**
  31  * Do a complex FFT with the parameters defined in ff_fft_init().
  32  * The input data must be permuted before with s->revtab table.
  33  * No 1.0 / sqrt(n) normalization is done.
  34  * AltiVec-enabled:
  35  * This code assumes that the 'z' pointer is 16 bytes-aligned.
  36  * It also assumes all FFTComplex are 8 bytes-aligned pairs of floats.
  37  */
  38
  39 void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z);
  40 void ff_fft_calc_interleave_altivec(FFTContext *s, FFTComplex *z);
  41
  42 #if HAVE_GNU_AS && HAVE_ALTIVEC && HAVE_BIGENDIAN
  43 static void imdct_half_altivec(FFTContext *s, FFTSample *output, const FFTSample *input)
  44 {
  45     int j, k;
  46     int n = 1 << s->mdct_bits;
  47     int n4 = n >> 2;
  48     int n8 = n >> 3;
  49     int n32 = n >> 5;
  50     const uint16_t *revtabj = s->revtab;
  51     const uint16_t *revtabk = s->revtab+n4;
  52     const vec_f *tcos = (const vec_f*)(s->tcos+n8);
  53     const vec_f *tsin = (const vec_f*)(s->tsin+n8);
  54     const vec_f *pin = (const vec_f*)(input+n4);
  55     vec_f *pout = (vec_f*)(output+n4);
  56
  57     /* pre rotation */
  58     k = n32-1;
  59     do {
  60         vec_f cos,sin,cos0,sin0,cos1,sin1,re,im,r0,i0,r1,i1,a,b,c,d;
  61 #define CMULA(p,o0,o1,o2,o3)\
  62         a = pin[ k*2+p];                       /* { z[k].re,    z[k].im,    z[k+1].re,  z[k+1].im  } */\
  63         b = pin[-k*2-p-1];                     /* { z[-k-2].re, z[-k-2].im, z[-k-1].re, z[-k-1].im } */\
  64         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 } */\
  65         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    } */\
  66         cos = vec_perm(cos0, cos1, vcprm(o0,o1,s##o2,s##o3)); /* { cos[k], cos[k+1], cos[-k-2], cos[-k-1] } */\
  67         sin = vec_perm(sin0, sin1, vcprm(o0,o1,s##o2,s##o3));\
  68         r##p = im*cos - re*sin;\
  69         i##p = re*cos + im*sin;
  70 #define STORE2(v,dst)\
  71         j = dst;\
  72         vec_ste(v, 0, output+j*2);\
  73         vec_ste(v, 4, output+j*2);
  74 #define STORE8(p)\
  75         a = vec_perm(r##p, i##p, vcprm(0,s0,0,s0));\
  76         b = vec_perm(r##p, i##p, vcprm(1,s1,1,s1));\
  77         c = vec_perm(r##p, i##p, vcprm(2,s2,2,s2));\
  78         d = vec_perm(r##p, i##p, vcprm(3,s3,3,s3));\
  79         STORE2(a, revtabk[ p*2-4]);\
  80         STORE2(b, revtabk[ p*2-3]);\
  81         STORE2(c, revtabj[-p*2+2]);\
  82         STORE2(d, revtabj[-p*2+3]);
  83
  84         cos0 = tcos[k];
  85         sin0 = tsin[k];
  86         cos1 = tcos[-k-1];
  87         sin1 = tsin[-k-1];
  88         CMULA(0, 0,1,2,3);
  89         CMULA(1, 2,3,0,1);
  90         STORE8(0);
  91         STORE8(1);
  92         revtabj += 4;
  93         revtabk -= 4;
  94         k--;
  95     } while(k >= 0);
  96
  97     ff_fft_calc_altivec(s, (FFTComplex*)output);
  98
  99     /* post rotation + reordering */
 100     j = -n32;
 101     k = n32-1;
 102     do {
 103         vec_f cos,sin,re,im,a,b,c,d;
 104 #define CMULB(d0,d1,o)\
 105         re = pout[o*2];\
 106         im = pout[o*2+1];\
 107         cos = tcos[o];\
 108         sin = tsin[o];\
 109         d0 = im*sin - re*cos;\
 110         d1 = re*sin + im*cos;
 111
 112         CMULB(a,b,j);
 113         CMULB(c,d,k);
 114         pout[2*j]   = vec_perm(a, d, vcprm(0,s3,1,s2));
 115         pout[2*j+1] = vec_perm(a, d, vcprm(2,s1,3,s0));
 116         pout[2*k]   = vec_perm(c, b, vcprm(0,s3,1,s2));
 117         pout[2*k+1] = vec_perm(c, b, vcprm(2,s1,3,s0));
 118         j++;
 119         k--;
 120     } while(k >= 0);
 121 }
 122
 123 static void imdct_calc_altivec(FFTContext *s, FFTSample *output, const FFTSample *input)
 124 {
 125     int k;
 126     int n = 1 << s->mdct_bits;
 127     int n4 = n >> 2;
 128     int n16 = n >> 4;
 129     vec_u32 sign = {1U<<31,1U<<31,1U<<31,1U<<31};
 130     vec_u32 *p0 = (vec_u32*)(output+n4);
 131     vec_u32 *p1 = (vec_u32*)(output+n4*3);
 132
 133     imdct_half_altivec(s, output + n4, input);
 134
 135     for (k = 0; k < n16; k++) {
 136         vec_u32 a = p0[k] ^ sign;
 137         vec_u32 b = p1[-k-1];
 138         p0[-k-1] = vec_perm(a, a, vcprm(3,2,1,0));
 139         p1[k]    = vec_perm(b, b, vcprm(3,2,1,0));
 140     }
 141 }
 142 #endif /* HAVE_GNU_AS && HAVE_ALTIVEC && HAVE_BIGENDIAN */
 143
 144 av_cold void ff_fft_init_ppc(FFTContext *s)
 145 {
 146 #if HAVE_GNU_AS && HAVE_ALTIVEC && HAVE_BIGENDIAN
 147     if (!PPC_ALTIVEC(av_get_cpu_flags()))
 148         return;
 149
 150     s->fft_calc   = ff_fft_calc_interleave_altivec;
 151     if (s->mdct_bits >= 5) {
 152         s->imdct_calc = imdct_calc_altivec;
 153         s->imdct_half = imdct_half_altivec;
 154     }
 155 #endif /* HAVE_GNU_AS && HAVE_ALTIVEC && HAVE_BIGENDIAN */
 156 }