libavcodec/i386/fft_3dn2.c

   1 /*
   2  * FFT/MDCT transform with Extended 3DNow! optimizations
   3  * Copyright (c) 2006 Zuxy MENG Jie, Loren Merritt
   4  * Based on fft_sse.c copyright (c) 2002 Fabrice Bellard.
   5  *
   6  * This file is part of FFmpeg.
   7  *
   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.
  12  *
  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.
  17  *
  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
  21  */
  22
  23 #include "libavutil/x86_cpu.h"
  24 #include "libavcodec/dsputil.h"
  25
  26 static const int p1m1[2] __attribute__((aligned(8))) =
  27     { 0, 1 << 31 };
  28
  29 static const int m1p1[2] __attribute__((aligned(8))) =
  30     { 1 << 31, 0 };
  31
  32 void ff_fft_calc_3dn2(FFTContext *s, FFTComplex *z)
  33 {
  34     int ln = s->nbits;
  35     long j;
  36     x86_reg i;
  37     long nblocks, nloops;
  38     FFTComplex *p, *cptr;
  39
  40     asm volatile(
  41         /* FEMMS is not a must here but recommended by AMD */
  42         "femms \n\t"
  43         "movq %0, %%mm7 \n\t"
  44         ::"m"(*(s->inverse ? m1p1 : p1m1))
  45     );
  46
  47     i = 8 << ln;
  48     asm volatile(
  49         "1: \n\t"
  50         "sub $32, %0 \n\t"
  51         "movq    (%0,%1), %%mm0 \n\t"
  52         "movq  16(%0,%1), %%mm1 \n\t"
  53         "movq   8(%0,%1), %%mm2 \n\t"
  54         "movq  24(%0,%1), %%mm3 \n\t"
  55         "movq      %%mm0, %%mm4 \n\t"
  56         "movq      %%mm1, %%mm5 \n\t"
  57         "pfadd     %%mm2, %%mm0 \n\t"
  58         "pfadd     %%mm3, %%mm1 \n\t"
  59         "pfsub     %%mm2, %%mm4 \n\t"
  60         "pfsub     %%mm3, %%mm5 \n\t"
  61         "movq      %%mm0, %%mm2 \n\t"
  62         "pswapd    %%mm5, %%mm5 \n\t"
  63         "movq      %%mm4, %%mm3 \n\t"
  64         "pxor      %%mm7, %%mm5 \n\t"
  65         "pfadd     %%mm1, %%mm0 \n\t"
  66         "pfadd     %%mm5, %%mm4 \n\t"
  67         "pfsub     %%mm1, %%mm2 \n\t"
  68         "pfsub     %%mm5, %%mm3 \n\t"
  69         "movq      %%mm0,   (%0,%1) \n\t"
  70         "movq      %%mm4,  8(%0,%1) \n\t"
  71         "movq      %%mm2, 16(%0,%1) \n\t"
  72         "movq      %%mm3, 24(%0,%1) \n\t"
  73         "jg 1b \n\t"
  74         :"+r"(i)
  75         :"r"(z)
  76     );
  77     /* pass 2 .. ln-1 */
  78
  79     nblocks = 1 << (ln-3);
  80     nloops = 1 << 2;
  81     cptr = s->exptab1;
  82     do {
  83         p = z;
  84         j = nblocks;
  85         do {
  86             i = nloops*8;
  87             asm volatile(
  88                 "1: \n\t"
  89                 "sub $16, %0 \n\t"
  90                 "movq    (%1,%0), %%mm0 \n\t"
  91                 "movq   8(%1,%0), %%mm1 \n\t"
  92                 "movq    (%2,%0), %%mm2 \n\t"
  93                 "movq   8(%2,%0), %%mm3 \n\t"
  94                 "movq  (%3,%0,2), %%mm4 \n\t"
  95                 "movq 8(%3,%0,2), %%mm5 \n\t"
  96                 "pswapd    %%mm4, %%mm6 \n\t" // no need for cptr[2] & cptr[3]
  97                 "pswapd    %%mm5, %%mm7 \n\t"
  98                 "pfmul     %%mm2, %%mm4 \n\t" // cre*re cim*im
  99                 "pfmul     %%mm3, %%mm5 \n\t"
 100                 "pfmul     %%mm2, %%mm6 \n\t" // cim*re cre*im
 101                 "pfmul     %%mm3, %%mm7 \n\t"
 102                 "pfpnacc   %%mm6, %%mm4 \n\t" // cre*re-cim*im cim*re+cre*im
 103                 "pfpnacc   %%mm7, %%mm5 \n\t"
 104                 "movq      %%mm0, %%mm2 \n\t"
 105                 "movq      %%mm1, %%mm3 \n\t"
 106                 "pfadd     %%mm4, %%mm0 \n\t"
 107                 "pfadd     %%mm5, %%mm1 \n\t"
 108                 "pfsub     %%mm4, %%mm2 \n\t"
 109                 "pfsub     %%mm5, %%mm3 \n\t"
 110                 "movq      %%mm0,  (%1,%0) \n\t"
 111                 "movq      %%mm1, 8(%1,%0) \n\t"
 112                 "movq      %%mm2,  (%2,%0) \n\t"
 113                 "movq      %%mm3, 8(%2,%0) \n\t"
 114                 "jg 1b \n\t"
 115                 :"+r"(i)
 116                 :"r"(p), "r"(p + nloops), "r"(cptr)
 117             );
 118             p += nloops*2;
 119         } while (--j);
 120         cptr += nloops*2;
 121         nblocks >>= 1;
 122         nloops <<= 1;
 123     } while (nblocks != 0);
 124     asm volatile("femms");
 125 }
 126
 127 void ff_imdct_calc_3dn2(MDCTContext *s, FFTSample *output,
 128                         const FFTSample *input, FFTSample *tmp)
 129 {
 130     long n8, n4, n2, n;
 131     x86_reg k;
 132     const uint16_t *revtab = s->fft.revtab;
 133     const FFTSample *tcos = s->tcos;
 134     const FFTSample *tsin = s->tsin;
 135     const FFTSample *in1, *in2;
 136     FFTComplex *z = (FFTComplex *)tmp;
 137
 138     n = 1 << s->nbits;
 139     n2 = n >> 1;
 140     n4 = n >> 2;
 141     n8 = n >> 3;
 142
 143     /* pre rotation */
 144     in1 = input;
 145     in2 = input + n2 - 1;
 146     for(k = 0; k < n4; k++) {
 147         // FIXME a single block is faster, but gcc 2.95 and 3.4.x on 32bit can't compile it
 148         asm volatile(
 149             "movd       %0, %%mm0 \n\t"
 150             "movd       %2, %%mm1 \n\t"
 151             "punpckldq  %1, %%mm0 \n\t"
 152             "punpckldq  %3, %%mm1 \n\t"
 153             "movq    %%mm0, %%mm2 \n\t"
 154             "pfmul   %%mm1, %%mm0 \n\t"
 155             "pswapd  %%mm1, %%mm1 \n\t"
 156             "pfmul   %%mm1, %%mm2 \n\t"
 157             "pfpnacc %%mm2, %%mm0 \n\t"
 158             ::"m"(in2[-2*k]), "m"(in1[2*k]),
 159               "m"(tcos[k]), "m"(tsin[k])
 160         );
 161         asm volatile(
 162             "movq    %%mm0, %0    \n\t"
 163             :"=m"(z[revtab[k]])
 164         );
 165     }
 166
 167     ff_fft_calc(&s->fft, z);
 168
 169     /* post rotation + reordering */
 170     for(k = 0; k < n4; k++) {
 171         asm volatile(
 172             "movq       %0, %%mm0 \n\t"
 173             "movd       %1, %%mm1 \n\t"
 174             "punpckldq  %2, %%mm1 \n\t"
 175             "movq    %%mm0, %%mm2 \n\t"
 176             "pfmul   %%mm1, %%mm0 \n\t"
 177             "pswapd  %%mm1, %%mm1 \n\t"
 178             "pfmul   %%mm1, %%mm2 \n\t"
 179             "pfpnacc %%mm2, %%mm0 \n\t"
 180             "movq    %%mm0, %0    \n\t"
 181             :"+m"(z[k])
 182             :"m"(tcos[k]), "m"(tsin[k])
 183         );
 184     }
 185
 186     k = n-8;
 187     asm volatile("movd %0, %%mm7" ::"r"(1<<31));
 188     asm volatile(
 189         "1: \n\t"
 190         "movq    (%4,%0), %%mm0 \n\t" // z[n8+k]
 191         "neg %0 \n\t"
 192         "pswapd -8(%4,%0), %%mm1 \n\t" // z[n8-1-k]
 193         "movq      %%mm0, %%mm2 \n\t"
 194         "pxor      %%mm7, %%mm2 \n\t"
 195         "punpckldq %%mm1, %%mm2 \n\t"
 196         "pswapd    %%mm2, %%mm3 \n\t"
 197         "punpckhdq %%mm1, %%mm0 \n\t"
 198         "pswapd    %%mm0, %%mm4 \n\t"
 199         "pxor      %%mm7, %%mm0 \n\t"
 200         "pxor      %%mm7, %%mm4 \n\t"
 201         "movq      %%mm3, -8(%3,%0) \n\t" // output[n-2-2*k] = { z[n8-1-k].im, -z[n8+k].re }
 202         "movq      %%mm4, -8(%2,%0) \n\t" // output[n2-2-2*k]= { -z[n8-1-k].re, z[n8+k].im }
 203         "neg %0 \n\t"
 204         "movq      %%mm0, (%1,%0) \n\t"   // output[2*k]     = { -z[n8+k].im, z[n8-1-k].re }
 205         "movq      %%mm2, (%2,%0) \n\t"   // output[n2+2*k]  = { -z[n8+k].re, z[n8-1-k].im }
 206         "sub $8, %0 \n\t"
 207         "jge 1b \n\t"
 208         :"+r"(k)
 209         :"r"(output), "r"(output+n2), "r"(output+n), "r"(z+n8)
 210         :"memory"
 211     );
 212     asm volatile("femms");
 213 }
 214