libavcodec/x86/fft_3dn2.c

   1 /*
   2  * FFT/MDCT transform with Extended 3DNow! optimizations
   3  * Copyright (c) 2006-2008 Zuxy MENG Jie, Loren Merritt
   4  *
   5  * This file is part of FFmpeg.
   6  *
   7  * FFmpeg is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2.1 of the License, or (at your option) any later version.
  11  *
  12  * FFmpeg is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with FFmpeg; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 #include "libavutil/x86_cpu.h"
  23 #include "libavcodec/dsputil.h"
  24
  25 DECLARE_ALIGNED_8(static const int, m1m1[2]) = { 1<<31, 1<<31 };
  26
  27 #ifdef EMULATE_3DNOWEXT
  28 #define PSWAPD(s,d)\
  29     "movq "#s","#d"\n"\
  30     "psrlq $32,"#d"\n"\
  31     "punpckldq "#s","#d"\n"
  32 #define ff_fft_calc_3dn2 ff_fft_calc_3dn
  33 #define ff_fft_dispatch_3dn2 ff_fft_dispatch_3dn
  34 #define ff_fft_dispatch_interleave_3dn2 ff_fft_dispatch_interleave_3dn
  35 #define ff_imdct_calc_3dn2 ff_imdct_calc_3dn
  36 #define ff_imdct_half_3dn2 ff_imdct_half_3dn
  37 #else
  38 #define PSWAPD(s,d) "pswapd "#s","#d"\n"
  39 #endif
  40
  41 void ff_fft_dispatch_3dn2(FFTComplex *z, int nbits);
  42 void ff_fft_dispatch_interleave_3dn2(FFTComplex *z, int nbits);
  43
  44 void ff_fft_calc_3dn2(FFTContext *s, FFTComplex *z)
  45 {
  46     int n = 1<<s->nbits;
  47     int i;
  48     ff_fft_dispatch_interleave_3dn2(z, s->nbits);
  49     __asm__ volatile("femms");
  50     if(n <= 8)
  51         for(i=0; i<n; i+=2)
  52             FFSWAP(FFTSample, z[i].im, z[i+1].re);
  53 }
  54
  55 void ff_imdct_half_3dn2(MDCTContext *s, FFTSample *output, const FFTSample *input)
  56 {
  57     x86_reg j, k;
  58     long n = 1 << s->nbits;
  59     long n2 = n >> 1;
  60     long n4 = n >> 2;
  61     long n8 = n >> 3;
  62     const uint16_t *revtab = s->fft.revtab;
  63     const FFTSample *tcos = s->tcos;
  64     const FFTSample *tsin = s->tsin;
  65     const FFTSample *in1, *in2;
  66     FFTComplex *z = (FFTComplex *)output;
  67
  68     /* pre rotation */
  69     in1 = input;
  70     in2 = input + n2 - 1;
  71 #ifdef EMULATE_3DNOWEXT
  72     __asm__ volatile("movd %0, %%mm7" ::"r"(1<<31));
  73 #endif
  74     for(k = 0; k < n4; k++) {
  75         // FIXME a single block is faster, but gcc 2.95 and 3.4.x on 32bit can't compile it
  76         __asm__ volatile(
  77             "movd         %0, %%mm0 \n"
  78             "movd         %2, %%mm1 \n"
  79             "punpckldq    %1, %%mm0 \n"
  80             "punpckldq    %3, %%mm1 \n"
  81             "movq      %%mm0, %%mm2 \n"
  82             PSWAPD(    %%mm1, %%mm3 )
  83             "pfmul     %%mm1, %%mm0 \n"
  84             "pfmul     %%mm3, %%mm2 \n"
  85 #ifdef EMULATE_3DNOWEXT
  86             "movq      %%mm0, %%mm1 \n"
  87             "punpckhdq %%mm2, %%mm0 \n"
  88             "punpckldq %%mm2, %%mm1 \n"
  89             "pxor      %%mm7, %%mm0 \n"
  90             "pfadd     %%mm1, %%mm0 \n"
  91 #else
  92             "pfpnacc   %%mm2, %%mm0 \n"
  93 #endif
  94             ::"m"(in2[-2*k]), "m"(in1[2*k]),
  95               "m"(tcos[k]), "m"(tsin[k])
  96         );
  97         __asm__ volatile(
  98             "movq    %%mm0, %0    \n\t"
  99             :"=m"(z[revtab[k]])
 100         );
 101     }
 102
 103     ff_fft_dispatch_3dn2(z, s->fft.nbits);
 104
 105 #define CMUL(j,mm0,mm1)\
 106         "movq  (%2,"#j",2), %%mm6 \n"\
 107         "movq 8(%2,"#j",2), "#mm0"\n"\
 108         "movq        %%mm6, "#mm1"\n"\
 109         "movq        "#mm0",%%mm7 \n"\
 110         "pfmul   (%3,"#j"), %%mm6 \n"\
 111         "pfmul   (%4,"#j"), "#mm0"\n"\
 112         "pfmul   (%4,"#j"), "#mm1"\n"\
 113         "pfmul   (%3,"#j"), %%mm7 \n"\
 114         "pfsub       %%mm6, "#mm0"\n"\
 115         "pfadd       %%mm7, "#mm1"\n"
 116
 117     /* post rotation */
 118     j = -n2;
 119     k = n2-8;
 120     __asm__ volatile(
 121         "1: \n"
 122         CMUL(%0, %%mm0, %%mm1)
 123         CMUL(%1, %%mm2, %%mm3)
 124         "movd   %%mm0,  (%2,%0,2) \n"
 125         "movd   %%mm1,12(%2,%1,2) \n"
 126         "movd   %%mm2,  (%2,%1,2) \n"
 127         "movd   %%mm3,12(%2,%0,2) \n"
 128         "psrlq  $32,   %%mm0 \n"
 129         "psrlq  $32,   %%mm1 \n"
 130         "psrlq  $32,   %%mm2 \n"
 131         "psrlq  $32,   %%mm3 \n"
 132         "movd   %%mm0, 8(%2,%0,2) \n"
 133         "movd   %%mm1, 4(%2,%1,2) \n"
 134         "movd   %%mm2, 8(%2,%1,2) \n"
 135         "movd   %%mm3, 4(%2,%0,2) \n"
 136         "sub $8, %1 \n"
 137         "add $8, %0 \n"
 138         "jl 1b \n"
 139         :"+r"(j), "+r"(k)
 140         :"r"(z+n8), "r"(tcos+n8), "r"(tsin+n8)
 141         :"memory"
 142     );
 143     __asm__ volatile("femms");
 144 }
 145
 146 void ff_imdct_calc_3dn2(MDCTContext *s, FFTSample *output, const FFTSample *input)
 147 {
 148     x86_reg j, k;
 149     long n = 1 << s->nbits;
 150     long n4 = n >> 2;
 151
 152     ff_imdct_half_3dn2(s, output+n4, input);
 153
 154     j = -n;
 155     k = n-8;
 156     __asm__ volatile(
 157         "movq %4, %%mm7 \n"
 158         "1: \n"
 159         PSWAPD((%2,%1), %%mm0)
 160         PSWAPD((%3,%0), %%mm1)
 161         "pxor    %%mm7, %%mm0 \n"
 162         "movq    %%mm1, (%3,%1) \n"
 163         "movq    %%mm0, (%2,%0) \n"
 164         "sub $8, %1 \n"
 165         "add $8, %0 \n"
 166         "jl 1b \n"
 167         :"+r"(j), "+r"(k)
 168         :"r"(output+n4), "r"(output+n4*3),
 169          "m"(*m1m1)
 170     );
 171     __asm__ volatile("femms");
 172 }
 173