libavcodec/ppc/fft_altivec.c

   1 /*
   2  * FFT/IFFT transforms
   3  * AltiVec-enabled
   4  * Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org>
   5  * Based on code Copyright (c) 2002 Fabrice Bellard.
   6  *
   7  * This file is part of FFmpeg.
   8  *
   9  * FFmpeg is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public
  11  * License as published by the Free Software Foundation; either
  12  * version 2.1 of the License, or (at your option) any later version.
  13  *
  14  * FFmpeg is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with FFmpeg; if not, write to the Free Software
  21  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  22  */
  23 #include "libavcodec/dsputil.h"
  24
  25 #include "gcc_fixes.h"
  26
  27 #include "dsputil_ppc.h"
  28 #include "util_altivec.h"
  29 /*
  30   those three macros are from libavcodec/fft.c
  31   and are required for the reference C code
  32 */
  33 /* butter fly op */
  34 #define BF(pre, pim, qre, qim, pre1, pim1, qre1, qim1) \
  35 {\
  36     FFTSample ax, ay, bx, by;\
  37     bx=pre1;\
  38     by=pim1;\
  39     ax=qre1;\
  40     ay=qim1;\
  41     pre = (bx + ax);\
  42     pim = (by + ay);\
  43     qre = (bx - ax);\
  44     qim = (by - ay);\
  45 }
  46 #define MUL16(a,b) ((a) * (b))
  47 #define CMUL(pre, pim, are, aim, bre, bim) \
  48 {\
  49     pre = (MUL16(are, bre) - MUL16(aim, bim));\
  50     pim = (MUL16(are, bim) + MUL16(bre, aim));\
  51 }
  52
  53
  54 /**
  55  * Do a complex FFT with the parameters defined in ff_fft_init(). The
  56  * input data must be permuted before with s->revtab table. No
  57  * 1.0/sqrt(n) normalization is done.
  58  * AltiVec-enabled
  59  * This code assumes that the 'z' pointer is 16 bytes-aligned
  60  * It also assumes all FFTComplex are 8 bytes-aligned pair of float
  61  * The code is exactly the same as the SSE version, except
  62  * that successive MUL + ADD/SUB have been merged into
  63  * fused multiply-add ('vec_madd' in altivec)
  64  */
  65 void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z)
  66 {
  67 POWERPC_PERF_DECLARE(altivec_fft_num, s->nbits >= 6);
  68     register const vector float vczero = (const vector float)vec_splat_u32(0.);
  69
  70     int ln = s->nbits;
  71     int j, np, np2;
  72     int nblocks, nloops;
  73     register FFTComplex *p, *q;
  74     FFTComplex *cptr, *cptr1;
  75     int k;
  76
  77 POWERPC_PERF_START_COUNT(altivec_fft_num, s->nbits >= 6);
  78
  79     np = 1 << ln;
  80
  81     {
  82         vector float *r, a, b, a1, c1, c2;
  83
  84         r = (vector float *)&z[0];
  85
  86         c1 = vcii(p,p,n,n);
  87
  88         if (s->inverse) {
  89             c2 = vcii(p,p,n,p);
  90         } else {
  91             c2 = vcii(p,p,p,n);
  92         }
  93
  94         j = (np >> 2);
  95         do {
  96             a = vec_ld(0, r);
  97             a1 = vec_ld(sizeof(vector float), r);
  98
  99             b = vec_perm(a,a,vcprmle(1,0,3,2));
 100             a = vec_madd(a,c1,b);
 101             /* do the pass 0 butterfly */
 102
 103             b = vec_perm(a1,a1,vcprmle(1,0,3,2));
 104             b = vec_madd(a1,c1,b);
 105             /* do the pass 0 butterfly */
 106
 107             /* multiply third by -i */
 108             b = vec_perm(b,b,vcprmle(2,3,1,0));
 109
 110             /* do the pass 1 butterfly */
 111             vec_st(vec_madd(b,c2,a), 0, r);
 112             vec_st(vec_nmsub(b,c2,a), sizeof(vector float), r);
 113
 114             r += 2;
 115         } while (--j != 0);
 116     }
 117     /* pass 2 .. ln-1 */
 118
 119     nblocks = np >> 3;
 120     nloops = 1 << 2;
 121     np2 = np >> 1;
 122
 123     cptr1 = s->exptab1;
 124     do {
 125         p = z;
 126         q = z + nloops;
 127         j = nblocks;
 128         do {
 129             cptr = cptr1;
 130             k = nloops >> 1;
 131             do {
 132                 vector float a,b,c,t1;
 133
 134                 a = vec_ld(0, (float*)p);
 135                 b = vec_ld(0, (float*)q);
 136
 137                 /* complex mul */
 138                 c = vec_ld(0, (float*)cptr);
 139                 /*  cre*re cim*re */
 140                 t1 = vec_madd(c, vec_perm(b,b,vcprmle(2,2,0,0)),vczero);
 141                 c = vec_ld(sizeof(vector float), (float*)cptr);
 142                 /*  -cim*im cre*im */
 143                 b = vec_madd(c, vec_perm(b,b,vcprmle(3,3,1,1)),t1);
 144
 145                 /* butterfly */
 146                 vec_st(vec_add(a,b), 0, (float*)p);
 147                 vec_st(vec_sub(a,b), 0, (float*)q);
 148
 149                 p += 2;
 150                 q += 2;
 151                 cptr += 4;
 152             } while (--k);
 153
 154             p += nloops;
 155             q += nloops;
 156         } while (--j);
 157         cptr1 += nloops * 2;
 158         nblocks = nblocks >> 1;
 159         nloops = nloops << 1;
 160     } while (nblocks != 0);
 161
 162 POWERPC_PERF_STOP_COUNT(altivec_fft_num, s->nbits >= 6);
 163 }