git.sesse.net Git - ffmpeg/blob - libavcodec/sh4/idct_sh4.c

   1 /*
   2  * idct for sh4
   3  *
   4  * Copyright (c) 2001-2003 BERO <bero@geocities.co.jp>
   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 "dsputil_sh4.h"
  24 #include "sh4.h"
  25
  26 #define c1      1.38703984532214752434  /* sqrt(2)*cos(1*pi/16) */
  27 #define c2      1.30656296487637657577  /* sqrt(2)*cos(2*pi/16) */
  28 #define c3      1.17587560241935884520  /* sqrt(2)*cos(3*pi/16) */
  29 #define c4      1.00000000000000000000  /* sqrt(2)*cos(4*pi/16) */
  30 #define c5      0.78569495838710234903  /* sqrt(2)*cos(5*pi/16) */
  31 #define c6      0.54119610014619712324  /* sqrt(2)*cos(6*pi/16) */
  32 #define c7      0.27589937928294311353  /* sqrt(2)*cos(7*pi/16) */
  33
  34 static const float even_table[] __attribute__ ((aligned(8))) = {
  35         c4, c4, c4, c4,
  36         c2, c6,-c6,-c2,
  37         c4,-c4,-c4, c4,
  38         c6,-c2, c2,-c6
  39 };
  40
  41 static const float odd_table[] __attribute__ ((aligned(8))) = {
  42         c1, c3, c5, c7,
  43         c3,-c7,-c1,-c5,
  44         c5,-c1, c7, c3,
  45         c7,-c5, c3,-c1
  46 };
  47
  48 #undef  c1
  49 #undef  c2
  50 #undef  c3
  51 #undef  c4
  52 #undef  c5
  53 #undef  c6
  54 #undef  c7
  55
  56 #define         load_matrix(table) \
  57     do { \
  58         const float *t = table; \
  59         __asm__ volatile( \
  60         "       fschg\n" \
  61         "       fmov   @%0+,xd0\n" \
  62         "       fmov   @%0+,xd2\n" \
  63         "       fmov   @%0+,xd4\n" \
  64         "       fmov   @%0+,xd6\n" \
  65         "       fmov   @%0+,xd8\n" \
  66         "       fmov   @%0+,xd10\n" \
  67         "       fmov   @%0+,xd12\n" \
  68         "       fmov   @%0+,xd14\n" \
  69         "       fschg\n" \
  70         : "+r"(t) \
  71         ); \
  72     } while (0)
  73
  74 #define         ftrv() \
  75                 __asm__ volatile("ftrv xmtrx,fv0" \
  76                 : "+f"(fr0),"+f"(fr1),"+f"(fr2),"+f"(fr3));
  77
  78 #define         DEFREG        \
  79         register float fr0 __asm__("fr0"); \
  80         register float fr1 __asm__("fr1"); \
  81         register float fr2 __asm__("fr2"); \
  82         register float fr3 __asm__("fr3")
  83
  84 #define         DESCALE(x,n)    (x)*(1.0f/(1<<(n)))
  85
  86 /* this code work worse on gcc cvs. 3.2.3 work fine */
  87
  88
  89 //optimized
  90
  91 void ff_idct_sh4(int16_t *block)
  92 {
  93         DEFREG;
  94
  95         int i;
  96         float        tblock[8*8],*fblock;
  97         int ofs1,ofs2,ofs3;
  98         int fpscr;
  99
 100         fp_single_enter(fpscr);
 101
 102         /* row */
 103
 104         /* even part */
 105         load_matrix(even_table);
 106
 107         fblock = tblock+4;
 108         i = 8;
 109         do {
 110                 fr0 = block[0];
 111                 fr1 = block[2];
 112                 fr2 = block[4];
 113                 fr3 = block[6];
 114                 block+=8;
 115                 ftrv();
 116                 *--fblock = fr3;
 117                 *--fblock = fr2;
 118                 *--fblock = fr1;
 119                 *--fblock = fr0;
 120                 fblock+=8+4;
 121         } while(--i);
 122         block-=8*8;
 123         fblock-=8*8+4;
 124
 125         load_matrix(odd_table);
 126
 127         i = 8;
 128
 129         do {
 130                 float t0,t1,t2,t3;
 131                 fr0 = block[1];
 132                 fr1 = block[3];
 133                 fr2 = block[5];
 134                 fr3 = block[7];
 135                 block+=8;
 136                 ftrv();
 137                 t0 = *fblock++;
 138                 t1 = *fblock++;
 139                 t2 = *fblock++;
 140                 t3 = *fblock++;
 141                 fblock+=4;
 142                 *--fblock = t0 - fr0;
 143                 *--fblock = t1 - fr1;
 144                 *--fblock = t2 - fr2;
 145                 *--fblock = t3 - fr3;
 146                 *--fblock = t3 + fr3;
 147                 *--fblock = t2 + fr2;
 148                 *--fblock = t1 + fr1;
 149                 *--fblock = t0 + fr0;
 150                 fblock+=8;
 151         } while(--i);
 152         block-=8*8;
 153         fblock-=8*8;
 154
 155         /* col */
 156
 157         /* even part */
 158         load_matrix(even_table);
 159
 160         ofs1 = sizeof(float)*2*8;
 161         ofs2 = sizeof(float)*4*8;
 162         ofs3 = sizeof(float)*6*8;
 163
 164         i = 8;
 165
 166 #define        OA(fblock,ofs)   *(float*)((char*)fblock + ofs)
 167
 168         do {
 169                 fr0 = OA(fblock,   0);
 170                 fr1 = OA(fblock,ofs1);
 171                 fr2 = OA(fblock,ofs2);
 172                 fr3 = OA(fblock,ofs3);
 173                 ftrv();
 174                 OA(fblock,0   ) = fr0;
 175                 OA(fblock,ofs1) = fr1;
 176                 OA(fblock,ofs2) = fr2;
 177                 OA(fblock,ofs3) = fr3;
 178                 fblock++;
 179         } while(--i);
 180         fblock-=8;
 181
 182         load_matrix(odd_table);
 183
 184         i=8;
 185         do {
 186                 float t0,t1,t2,t3;
 187                 t0 = OA(fblock,   0); /* [8*0] */
 188                 t1 = OA(fblock,ofs1); /* [8*2] */
 189                 t2 = OA(fblock,ofs2); /* [8*4] */
 190                 t3 = OA(fblock,ofs3); /* [8*6] */
 191                 fblock+=8;
 192                 fr0 = OA(fblock,   0); /* [8*1] */
 193                 fr1 = OA(fblock,ofs1); /* [8*3] */
 194                 fr2 = OA(fblock,ofs2); /* [8*5] */
 195                 fr3 = OA(fblock,ofs3); /* [8*7] */
 196                 fblock+=-8+1;
 197                 ftrv();
 198                 block[8*0] = DESCALE(t0 + fr0,3);
 199                 block[8*7] = DESCALE(t0 - fr0,3);
 200                 block[8*1] = DESCALE(t1 + fr1,3);
 201                 block[8*6] = DESCALE(t1 - fr1,3);
 202                 block[8*2] = DESCALE(t2 + fr2,3);
 203                 block[8*5] = DESCALE(t2 - fr2,3);
 204                 block[8*3] = DESCALE(t3 + fr3,3);
 205                 block[8*4] = DESCALE(t3 - fr3,3);
 206                 block++;
 207         } while(--i);
 208
 209         fp_single_leave(fpscr);
 210 }