libavcodec/i386/vp3dsp_sse2.c

   1 /*
   2  * Copyright (C) 2004 the ffmpeg project
   3  *
   4  * This file is part of FFmpeg.
   5  *
   6  * FFmpeg is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2.1 of the License, or (at your option) any later version.
  10  *
  11  * FFmpeg is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with FFmpeg; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  19  */
  20
  21 /**
  22  * @file vp3dsp_sse2.c
  23  * SSE2-optimized functions cribbed from the original VP3 source code.
  24  */
  25
  26 #include "libavcodec/dsputil.h"
  27 #include "mmx.h"
  28
  29 static DECLARE_ALIGNED_16(const unsigned short, SSE2_dequant_const[]) =
  30 {
  31     0,65535,65535,0,0,0,0,0,    // 0x0000 0000 0000 0000 0000 FFFF FFFF 0000
  32     0,0,0,0,65535,65535,0,0,    // 0x0000 0000 FFFF FFFF 0000 0000 0000 0000
  33     65535,65535,65535,0,0,0,0,0,// 0x0000 0000 0000 0000 0000 FFFF FFFF FFFF
  34     0,0,0,65535,0,0,0,0,        // 0x0000 0000 0000 0000 FFFF 0000 0000 0000
  35     0,0,0,65535,65535,0,0,0,    // 0x0000 0000 0000 FFFF FFFF 0000 0000 0000
  36     65535,0,0,0,0,65535,0,0,    // 0x0000 0000 FFFF 0000 0000 0000 0000 FFFF
  37     0,0,65535,65535, 0,0,0,0    // 0x0000 0000 0000 0000 FFFF FFFF 0000 0000
  38 };
  39
  40 static DECLARE_ALIGNED_16(const unsigned int, eight_data[]) =
  41 {
  42     0x00080008,
  43     0x00080008,
  44     0x00080008,
  45     0x00080008
  46 };
  47
  48 static DECLARE_ALIGNED_16(const unsigned short, SSE2_idct_data[7 * 8]) =
  49 {
  50     64277,64277,64277,64277,64277,64277,64277,64277,
  51     60547,60547,60547,60547,60547,60547,60547,60547,
  52     54491,54491,54491,54491,54491,54491,54491,54491,
  53     46341,46341,46341,46341,46341,46341,46341,46341,
  54     36410,36410,36410,36410,36410,36410,36410,36410,
  55     25080,25080,25080,25080,25080,25080,25080,25080,
  56     12785,12785,12785,12785,12785,12785,12785,12785
  57 };
  58
  59
  60 #define SSE2_Column_IDCT() {        \
  61     \
  62     movdqa_m2r(*I(3), xmm2);     /* xmm2 = i3 */ \
  63     movdqa_m2r(*C(3), xmm6);     /* xmm6 = c3 */ \
  64     \
  65     movdqa_r2r(xmm2, xmm4);      /* xmm4 = i3 */ \
  66     movdqa_m2r(*I(5), xmm7);     /* xmm7 = i5 */ \
  67     \
  68     pmulhw_r2r(xmm6, xmm4);      /* xmm4 = c3 * i3 - i3 */ \
  69     movdqa_m2r(*C(5), xmm1);     /* xmm1 = c5 */ \
  70     \
  71     pmulhw_r2r(xmm7, xmm6);      /* xmm6 = c3 * i5 - i5 */ \
  72     movdqa_r2r(xmm1, xmm5);      /* xmm5 = c5 */ \
  73     \
  74     pmulhw_r2r(xmm2, xmm1);      /* xmm1 = c5 * i3 - i3 */ \
  75     movdqa_m2r(*I(1), xmm3);     /* xmm3 = i1 */ \
  76     \
  77     pmulhw_r2r(xmm7, xmm5);      /* xmm5 = c5 * i5 - i5 */ \
  78     movdqa_m2r(*C(1), xmm0);     /* xmm0 = c1 */ \
  79     \
  80     /* all registers are in use */ \
  81     \
  82     paddw_r2r(xmm2, xmm4);       /* xmm4 = c3 * i3 */ \
  83     paddw_r2r(xmm7, xmm6);       /* xmm6 = c3 * i5 */ \
  84     \
  85     paddw_r2r(xmm1, xmm2);       /* xmm2 = c5 * i3 */ \
  86     movdqa_m2r(*I(7), xmm1);     /* xmm1 = i7 */ \
  87     \
  88     paddw_r2r(xmm5, xmm7);       /* xmm7 = c5 * i5 */ \
  89     movdqa_r2r(xmm0, xmm5);      /* xmm5 = c1 */ \
  90     \
  91     pmulhw_r2r(xmm3, xmm0);      /* xmm0 = c1 * i1 - i1 */ \
  92     paddsw_r2r(xmm7, xmm4);      /* xmm4 = c3 * i3 + c5 * i5 = C */ \
  93     \
  94     pmulhw_r2r(xmm1, xmm5);      /* xmm5 = c1 * i7 - i7 */ \
  95     movdqa_m2r(*C(7), xmm7);     /* xmm7 = c7 */ \
  96     \
  97     psubsw_r2r(xmm2, xmm6);      /* xmm6 = c3 * i5 - c5 * i3 = D */ \
  98     paddw_r2r(xmm3, xmm0);       /* xmm0 = c1 * i1 */ \
  99     \
 100     pmulhw_r2r(xmm7, xmm3);      /* xmm3 = c7 * i1 */ \
 101     movdqa_m2r(*I(2), xmm2);     /* xmm2 = i2 */ \
 102     \
 103     pmulhw_r2r(xmm1, xmm7);      /* xmm7 = c7 * i7 */ \
 104     paddw_r2r(xmm1, xmm5);       /* xmm5 = c1 * i7 */ \
 105     \
 106     movdqa_r2r(xmm2, xmm1);      /* xmm1 = i2 */ \
 107     pmulhw_m2r(*C(2), xmm2);     /* xmm2 = i2 * c2 -i2 */ \
 108     \
 109     psubsw_r2r(xmm5, xmm3);      /* xmm3 = c7 * i1 - c1 * i7 = B */ \
 110     movdqa_m2r(*I(6), xmm5);     /* xmm5 = i6 */ \
 111     \
 112     paddsw_r2r(xmm7, xmm0);      /* xmm0 = c1 * i1 + c7 * i7 = A */ \
 113     movdqa_r2r(xmm5, xmm7);      /* xmm7 = i6 */ \
 114     \
 115     psubsw_r2r(xmm4, xmm0);      /* xmm0 = A - C */ \
 116     pmulhw_m2r(*C(2), xmm5);     /* xmm5 = c2 * i6 - i6 */ \
 117     \
 118     paddw_r2r(xmm1, xmm2);       /* xmm2 = i2 * c2 */ \
 119     pmulhw_m2r(*C(6), xmm1);     /* xmm1 = c6 * i2 */ \
 120     \
 121     paddsw_r2r(xmm4, xmm4);      /* xmm4 = C + C */ \
 122     paddsw_r2r(xmm0, xmm4);      /* xmm4 = A + C = C. */ \
 123     \
 124     psubsw_r2r(xmm6, xmm3);      /* xmm3 = B - D */ \
 125     paddw_r2r(xmm7, xmm5);       /* xmm5 = c2 * i6 */ \
 126     \
 127     paddsw_r2r(xmm6, xmm6);      /* xmm6 = D + D */ \
 128     pmulhw_m2r(*C(6), xmm7);     /* xmm7 = c6 * i6 */ \
 129     \
 130     paddsw_r2r(xmm3, xmm6);      /* xmm6 = B + D = D. */ \
 131     movdqa_r2m(xmm4, *I(1));     /* Save C. at I(1) */ \
 132     \
 133     psubsw_r2r(xmm5, xmm1);      /* xmm1 = c6 * i2 - c2 * i6 = H */ \
 134     movdqa_m2r(*C(4), xmm4);     /* xmm4 = c4 */ \
 135     \
 136     movdqa_r2r(xmm3, xmm5);      /* xmm5 = B - D */ \
 137     pmulhw_r2r(xmm4, xmm3);      /* xmm3 = ( c4 -1 ) * ( B - D ) */ \
 138     \
 139     paddsw_r2r(xmm2, xmm7);      /* xmm7 = c2 * i2 + c6 * i6 = G */ \
 140     movdqa_r2m(xmm6, *I(2));     /* Save D. at I(2) */ \
 141     \
 142     movdqa_r2r(xmm0, xmm2);      /* xmm2 = A - C */ \
 143     movdqa_m2r(*I(0), xmm6);     /* xmm6 = i0 */ \
 144     \
 145     pmulhw_r2r(xmm4, xmm0);      /* xmm0 = ( c4 - 1 ) * ( A - C ) = A. */ \
 146     paddw_r2r(xmm3, xmm5);       /* xmm5 = c4 * ( B - D ) = B. */ \
 147     \
 148     movdqa_m2r(*I(4), xmm3);     /* xmm3 = i4 */ \
 149     psubsw_r2r(xmm1, xmm5);      /* xmm5 = B. - H = B.. */ \
 150     \
 151     paddw_r2r(xmm0, xmm2);       /* xmm2 = c4 * ( A - C) = A. */ \
 152     psubsw_r2r(xmm3, xmm6);      /* xmm6 = i0 - i4 */ \
 153     \
 154     movdqa_r2r(xmm6, xmm0);      /* xmm0 = i0 - i4 */ \
 155     pmulhw_r2r(xmm4, xmm6);      /* xmm6 = (c4 - 1) * (i0 - i4) = F */ \
 156     \
 157     paddsw_r2r(xmm3, xmm3);      /* xmm3 = i4 + i4 */ \
 158     paddsw_r2r(xmm1, xmm1);      /* xmm1 = H + H */ \
 159     \
 160     paddsw_r2r(xmm0, xmm3);      /* xmm3 = i0 + i4 */ \
 161     paddsw_r2r(xmm5, xmm1);      /* xmm1 = B. + H = H. */ \
 162     \
 163     pmulhw_r2r(xmm3, xmm4);      /* xmm4 = ( c4 - 1 ) * ( i0 + i4 )  */ \
 164     paddw_r2r(xmm0, xmm6);       /* xmm6 = c4 * ( i0 - i4 ) */ \
 165     \
 166     psubsw_r2r(xmm2, xmm6);      /* xmm6 = F - A. = F. */ \
 167     paddsw_r2r(xmm2, xmm2);      /* xmm2 = A. + A. */ \
 168     \
 169     movdqa_m2r(*I(1), xmm0);     /* Load        C. from I(1) */ \
 170     paddsw_r2r(xmm6, xmm2);      /* xmm2 = F + A. = A.. */ \
 171     \
 172     paddw_r2r(xmm3, xmm4);       /* xmm4 = c4 * ( i0 + i4 ) = 3 */ \
 173     psubsw_r2r(xmm1, xmm2);      /* xmm2 = A.. - H. = R2 */ \
 174     \
 175     paddsw_m2r(*Eight, xmm2);    /* Adjust R2 and R1 before shifting */ \
 176     paddsw_r2r(xmm1, xmm1);      /* xmm1 = H. + H. */ \
 177     \
 178     paddsw_r2r(xmm2, xmm1);      /* xmm1 = A.. + H. = R1 */ \
 179     psraw_i2r(4, xmm2);          /* xmm2 = op2 */ \
 180     \
 181     psubsw_r2r(xmm7, xmm4);      /* xmm4 = E - G = E. */ \
 182     psraw_i2r(4, xmm1);          /* xmm1 = op1 */ \
 183     \
 184     movdqa_m2r(*I(2), xmm3);     /* Load D. from I(2) */ \
 185     paddsw_r2r(xmm7, xmm7);      /* xmm7 = G + G */ \
 186     \
 187     movdqa_r2m(xmm2, *O(2));     /* Write out op2 */ \
 188     paddsw_r2r(xmm4, xmm7);      /* xmm7 = E + G = G. */ \
 189     \
 190     movdqa_r2m(xmm1, *O(1));     /* Write out op1 */ \
 191     psubsw_r2r(xmm3, xmm4);      /* xmm4 = E. - D. = R4 */ \
 192     \
 193     paddsw_m2r(*Eight, xmm4);    /* Adjust R4 and R3 before shifting */ \
 194     paddsw_r2r(xmm3, xmm3);      /* xmm3 = D. + D. */ \
 195     \
 196     paddsw_r2r(xmm4, xmm3);      /* xmm3 = E. + D. = R3 */ \
 197     psraw_i2r(4, xmm4);          /* xmm4 = op4 */ \
 198     \
 199     psubsw_r2r(xmm5, xmm6);      /* xmm6 = F. - B..= R6 */ \
 200     psraw_i2r(4, xmm3);          /* xmm3 = op3 */ \
 201     \
 202     paddsw_m2r(*Eight, xmm6);    /* Adjust R6 and R5 before shifting */ \
 203     paddsw_r2r(xmm5, xmm5);      /* xmm5 = B.. + B.. */ \
 204     \
 205     paddsw_r2r(xmm6, xmm5);      /* xmm5 = F. + B.. = R5 */ \
 206     psraw_i2r(4, xmm6);          /* xmm6 = op6 */ \
 207     \
 208     movdqa_r2m(xmm4, *O(4));     /* Write out op4 */ \
 209     psraw_i2r(4, xmm5);          /* xmm5 = op5 */ \
 210     \
 211     movdqa_r2m(xmm3, *O(3));     /* Write out op3 */ \
 212     psubsw_r2r(xmm0, xmm7);      /* xmm7 = G. - C. = R7 */ \
 213     \
 214     paddsw_m2r(*Eight, xmm7);    /* Adjust R7 and R0 before shifting */ \
 215     paddsw_r2r(xmm0, xmm0);      /* xmm0 = C. + C. */ \
 216     \
 217     paddsw_r2r(xmm7, xmm0);      /* xmm0 = G. + C. */ \
 218     psraw_i2r(4, xmm7);          /* xmm7 = op7 */ \
 219     \
 220     movdqa_r2m(xmm6, *O(6));     /* Write out op6 */ \
 221     psraw_i2r(4, xmm0);          /* xmm0 = op0 */ \
 222     \
 223     movdqa_r2m(xmm5, *O(5));     /* Write out op5 */ \
 224     movdqa_r2m(xmm7, *O(7));     /* Write out op7 */ \
 225     \
 226     movdqa_r2m(xmm0, *O(0));     /* Write out op0 */ \
 227     \
 228 } /* End of SSE2_Column_IDCT macro */
 229
 230
 231 #define SSE2_Row_IDCT() {        \
 232     \
 233     movdqa_m2r(*I(3), xmm2);     /* xmm2 = i3 */ \
 234     movdqa_m2r(*C(3), xmm6);     /* xmm6 = c3 */ \
 235     \
 236     movdqa_r2r(xmm2, xmm4);      /* xmm4 = i3 */ \
 237     movdqa_m2r(*I(5), xmm7);     /* xmm7 = i5 */ \
 238     \
 239     pmulhw_r2r(xmm6, xmm4);      /* xmm4 = c3 * i3 - i3 */ \
 240     movdqa_m2r(*C(5), xmm1);     /* xmm1 = c5 */ \
 241     \
 242     pmulhw_r2r(xmm7, xmm6);      /* xmm6 = c3 * i5 - i5 */ \
 243     movdqa_r2r(xmm1, xmm5);      /* xmm5 = c5 */ \
 244     \
 245     pmulhw_r2r(xmm2, xmm1);      /* xmm1 = c5 * i3 - i3 */ \
 246     movdqa_m2r(*I(1), xmm3);     /* xmm3 = i1 */ \
 247     \
 248     pmulhw_r2r(xmm7, xmm5);      /* xmm5 = c5 * i5 - i5 */ \
 249     movdqa_m2r(*C(1), xmm0);     /* xmm0 = c1 */ \
 250     \
 251     /* all registers are in use */ \
 252     \
 253     paddw_r2r(xmm2, xmm4);      /* xmm4 = c3 * i3 */ \
 254     paddw_r2r(xmm7, xmm6);      /* xmm6 = c3 * i5 */ \
 255     \
 256     paddw_r2r(xmm1, xmm2);      /* xmm2 = c5 * i3 */ \
 257     movdqa_m2r(*I(7), xmm1);    /* xmm1 = i7 */ \
 258     \
 259     paddw_r2r(xmm5, xmm7);      /* xmm7 = c5 * i5 */ \
 260     movdqa_r2r(xmm0, xmm5);     /* xmm5 = c1 */ \
 261     \
 262     pmulhw_r2r(xmm3, xmm0);     /* xmm0 = c1 * i1 - i1 */ \
 263     paddsw_r2r(xmm7, xmm4);     /* xmm4 = c3 * i3 + c5 * i5 = C */ \
 264     \
 265     pmulhw_r2r(xmm1, xmm5);     /* xmm5 = c1 * i7 - i7 */ \
 266     movdqa_m2r(*C(7), xmm7);    /* xmm7 = c7 */ \
 267     \
 268     psubsw_r2r(xmm2, xmm6);     /* xmm6 = c3 * i5 - c5 * i3 = D */ \
 269     paddw_r2r(xmm3, xmm0);      /* xmm0 = c1 * i1 */ \
 270     \
 271     pmulhw_r2r(xmm7, xmm3);     /* xmm3 = c7 * i1 */ \
 272     movdqa_m2r(*I(2), xmm2);    /* xmm2 = i2 */ \
 273     \
 274     pmulhw_r2r(xmm1, xmm7);     /* xmm7 = c7 * i7 */ \
 275     paddw_r2r(xmm1, xmm5);      /* xmm5 = c1 * i7 */ \
 276     \
 277     movdqa_r2r(xmm2, xmm1);     /* xmm1 = i2 */ \
 278     pmulhw_m2r(*C(2), xmm2);    /* xmm2 = i2 * c2 -i2 */ \
 279     \
 280     psubsw_r2r(xmm5, xmm3);     /* xmm3 = c7 * i1 - c1 * i7 = B */ \
 281     movdqa_m2r(*I(6), xmm5);    /* xmm5 = i6 */ \
 282     \
 283     paddsw_r2r(xmm7, xmm0);     /* xmm0 = c1 * i1 + c7 * i7        = A */ \
 284     movdqa_r2r(xmm5, xmm7);     /* xmm7 = i6 */ \
 285     \
 286     psubsw_r2r(xmm4, xmm0);     /* xmm0 = A - C */ \
 287     pmulhw_m2r(*C(2), xmm5);    /* xmm5 = c2 * i6 - i6 */ \
 288     \
 289     paddw_r2r(xmm1, xmm2);      /* xmm2 = i2 * c2 */ \
 290     pmulhw_m2r(*C(6), xmm1);    /* xmm1 = c6 * i2 */ \
 291     \
 292     paddsw_r2r(xmm4, xmm4);     /* xmm4 = C + C */ \
 293     paddsw_r2r(xmm0, xmm4);     /* xmm4 = A + C = C. */ \
 294     \
 295     psubsw_r2r(xmm6, xmm3);     /* xmm3 = B - D */ \
 296     paddw_r2r(xmm7, xmm5);      /* xmm5 = c2 * i6 */ \
 297     \
 298     paddsw_r2r(xmm6, xmm6);     /* xmm6 = D + D */ \
 299     pmulhw_m2r(*C(6), xmm7);    /* xmm7 = c6 * i6 */ \
 300     \
 301     paddsw_r2r(xmm3, xmm6);     /* xmm6 = B + D = D. */ \
 302     movdqa_r2m(xmm4, *I(1));    /* Save C. at I(1)        */ \
 303     \
 304     psubsw_r2r(xmm5, xmm1);     /* xmm1 = c6 * i2 - c2 * i6 = H */ \
 305     movdqa_m2r(*C(4), xmm4);    /* xmm4 = c4 */ \
 306     \
 307     movdqa_r2r(xmm3, xmm5);     /* xmm5 = B - D */ \
 308     pmulhw_r2r(xmm4, xmm3);     /* xmm3 = ( c4 -1 ) * ( B - D ) */ \
 309     \
 310     paddsw_r2r(xmm2, xmm7);     /* xmm7 = c2 * i2 + c6 * i6 = G */ \
 311     movdqa_r2m(xmm6, *I(2));    /* Save D. at I(2) */ \
 312     \
 313     movdqa_r2r(xmm0, xmm2);     /* xmm2 = A - C */ \
 314     movdqa_m2r(*I(0), xmm6);    /* xmm6 = i0 */ \
 315     \
 316     pmulhw_r2r(xmm4, xmm0);     /* xmm0 = ( c4 - 1 ) * ( A - C ) = A. */ \
 317     paddw_r2r(xmm3, xmm5);      /* xmm5 = c4 * ( B - D ) = B. */ \
 318     \
 319     movdqa_m2r(*I(4), xmm3);    /* xmm3 = i4 */ \
 320     psubsw_r2r(xmm1, xmm5);     /* xmm5 = B. - H = B.. */ \
 321     \
 322     paddw_r2r(xmm0, xmm2);      /* xmm2 = c4 * ( A - C) = A. */ \
 323     psubsw_r2r(xmm3, xmm6);     /* xmm6 = i0 - i4 */ \
 324     \
 325     movdqa_r2r(xmm6, xmm0);     /* xmm0 = i0 - i4 */ \
 326     pmulhw_r2r(xmm4, xmm6);     /* xmm6 = ( c4 - 1 ) * ( i0 - i4 ) = F */ \
 327     \
 328     paddsw_r2r(xmm3, xmm3);     /* xmm3 = i4 + i4 */ \
 329     paddsw_r2r(xmm1, xmm1);     /* xmm1 = H + H */ \
 330     \
 331     paddsw_r2r(xmm0, xmm3);     /* xmm3 = i0 + i4 */ \
 332     paddsw_r2r(xmm5, xmm1);     /* xmm1 = B. + H = H. */ \
 333     \
 334     pmulhw_r2r(xmm3, xmm4);     /* xmm4 = ( c4 - 1 ) * ( i0 + i4 )  */ \
 335     paddw_r2r(xmm0, xmm6);      /* xmm6 = c4 * ( i0 - i4 ) */ \
 336     \
 337     psubsw_r2r(xmm2, xmm6);     /* xmm6 = F - A. = F. */ \
 338     paddsw_r2r(xmm2, xmm2);     /* xmm2 = A. + A. */ \
 339     \
 340     movdqa_m2r(*I(1), xmm0);    /* Load C. from I(1) */ \
 341     paddsw_r2r(xmm6, xmm2);     /* xmm2 = F + A. = A.. */ \
 342     \
 343     paddw_r2r(xmm3, xmm4);      /* xmm4 = c4 * ( i0 + i4 ) = 3 */ \
 344     psubsw_r2r(xmm1, xmm2);     /* xmm2 = A.. - H. = R2 */ \
 345     \
 346     paddsw_r2r(xmm1, xmm1);     /* xmm1 = H. + H. */ \
 347     paddsw_r2r(xmm2, xmm1);     /* xmm1 = A.. + H. = R1 */ \
 348     \
 349     psubsw_r2r(xmm7, xmm4);     /* xmm4 = E - G = E. */ \
 350     \
 351     movdqa_m2r(*I(2), xmm3);    /* Load D. from I(2) */ \
 352     paddsw_r2r(xmm7, xmm7);     /* xmm7 = G + G */ \
 353     \
 354     movdqa_r2m(xmm2, *I(2));    /* Write out op2 */ \
 355     paddsw_r2r(xmm4, xmm7);     /* xmm7 = E + G = G. */ \
 356     \
 357     movdqa_r2m(xmm1, *I(1));    /* Write out op1 */ \
 358     psubsw_r2r(xmm3, xmm4);     /* xmm4 = E. - D. = R4 */ \
 359     \
 360     paddsw_r2r(xmm3, xmm3);     /* xmm3 = D. + D. */ \
 361     \
 362     paddsw_r2r(xmm4, xmm3);     /* xmm3 = E. + D. = R3 */ \
 363     \
 364     psubsw_r2r(xmm5, xmm6);     /* xmm6 = F. - B..= R6 */ \
 365     \
 366     paddsw_r2r(xmm5, xmm5);     /* xmm5 = B.. + B.. */ \
 367     \
 368     paddsw_r2r(xmm6, xmm5);     /* xmm5 = F. + B.. = R5 */ \
 369     \
 370     movdqa_r2m(xmm4, *I(4));    /* Write out op4 */ \
 371     \
 372     movdqa_r2m(xmm3, *I(3));    /* Write out op3 */ \
 373     psubsw_r2r(xmm0, xmm7);     /* xmm7 = G. - C. = R7 */ \
 374     \
 375     paddsw_r2r(xmm0, xmm0);     /* xmm0 = C. + C. */ \
 376     \
 377     paddsw_r2r(xmm7, xmm0);     /* xmm0 = G. + C. */ \
 378     \
 379     movdqa_r2m(xmm6, *I(6));    /* Write out op6 */ \
 380     \
 381     movdqa_r2m(xmm5, *I(5));    /* Write out op5 */ \
 382     movdqa_r2m(xmm7, *I(7));    /* Write out op7 */ \
 383     \
 384     movdqa_r2m(xmm0, *I(0));    /* Write out op0 */ \
 385     \
 386 } /* End of SSE2_Row_IDCT macro */
 387
 388
 389 #define SSE2_Transpose() {    \
 390     \
 391     movdqa_m2r(*I(4), xmm4);    /* xmm4=e7e6e5e4e3e2e1e0 */ \
 392     movdqa_m2r(*I(5), xmm0);    /* xmm4=f7f6f5f4f3f2f1f0 */ \
 393     \
 394     movdqa_r2r(xmm4, xmm5);     /* make a copy */ \
 395     punpcklwd_r2r(xmm0, xmm4);  /* xmm4=f3e3f2e2f1e1f0e0 */ \
 396     \
 397     punpckhwd_r2r(xmm0, xmm5);  /* xmm5=f7e7f6e6f5e5f4e4 */ \
 398     movdqa_m2r(*I(6), xmm6);    /* xmm6=g7g6g5g4g3g2g1g0 */ \
 399     \
 400     movdqa_m2r(*I(7), xmm0);    /* xmm0=h7h6h5h4h3h2h1h0 */ \
 401     movdqa_r2r(xmm6, xmm7);     /* make a copy */ \
 402     \
 403     punpcklwd_r2r(xmm0, xmm6);  /* xmm6=h3g3h3g2h1g1h0g0 */ \
 404     punpckhwd_r2r(xmm0, xmm7);  /* xmm7=h7g7h6g6h5g5h4g4 */ \
 405     \
 406     movdqa_r2r(xmm4, xmm3);     /* make a copy */ \
 407     punpckldq_r2r(xmm6, xmm4);  /* xmm4=h1g1f1e1h0g0f0e0 */ \
 408     \
 409     punpckhdq_r2r(xmm6, xmm3);  /* xmm3=h3g3g3e3h2g2f2e2 */ \
 410     movdqa_r2m(xmm3, *I(6));    /* save h3g3g3e3h2g2f2e2 */ \
 411     /* Free xmm6 */ \
 412     movdqa_r2r(xmm5, xmm6);     /* make a copy */ \
 413     punpckldq_r2r(xmm7, xmm5);  /* xmm5=h5g5f5e5h4g4f4e4 */ \
 414     \
 415     punpckhdq_r2r(xmm7, xmm6);  /* xmm6=h7g7f7e7h6g6f6e6 */ \
 416     movdqa_m2r(*I(0), xmm0);    /* xmm0=a7a6a5a4a3a2a1a0 */ \
 417     /* Free xmm7 */ \
 418     movdqa_m2r(*I(1), xmm1);    /* xmm1=b7b6b5b4b3b2b1b0 */ \
 419     movdqa_r2r(xmm0, xmm7);     /* make a copy */ \
 420     \
 421     punpcklwd_r2r(xmm1, xmm0);  /* xmm0=b3a3b2a2b1a1b0a0 */ \
 422     punpckhwd_r2r(xmm1, xmm7);  /* xmm7=b7a7b6a6b5a5b4a4 */ \
 423     /* Free xmm1 */ \
 424     movdqa_m2r(*I(2), xmm2);    /* xmm2=c7c6c5c4c3c2c1c0 */ \
 425     movdqa_m2r(*I(3), xmm3);    /* xmm3=d7d6d5d4d3d2d1d0 */ \
 426     \
 427     movdqa_r2r(xmm2, xmm1);     /* make a copy */ \
 428     punpcklwd_r2r(xmm3, xmm2);  /* xmm2=d3c3d2c2d1c1d0c0 */ \
 429     \
 430     punpckhwd_r2r(xmm3, xmm1);  /* xmm1=d7c7d6c6d5c5d4c4 */ \
 431     movdqa_r2r(xmm0, xmm3);     /* make a copy        */ \
 432     \
 433     punpckldq_r2r(xmm2, xmm0);  /* xmm0=d1c1b1a1d0c0b0a0 */ \
 434     punpckhdq_r2r(xmm2, xmm3);  /* xmm3=d3c3b3a3d2c2b2a2 */ \
 435     /* Free xmm2 */ \
 436     movdqa_r2r(xmm7, xmm2);     /* make a copy */ \
 437     punpckldq_r2r(xmm1, xmm2);  /* xmm2=d5c5b5a5d4c4b4a4 */ \
 438     \
 439     punpckhdq_r2r(xmm1, xmm7);  /* xmm7=d7c7b7a7d6c6b6a6 */ \
 440     movdqa_r2r(xmm0, xmm1);     /* make a copy */ \
 441     \
 442     punpcklqdq_r2r(xmm4, xmm0); /* xmm0=h0g0f0e0d0c0b0a0 */ \
 443     punpckhqdq_r2r(xmm4, xmm1); /* xmm1=h1g1g1e1d1c1b1a1 */ \
 444     \
 445     movdqa_r2m(xmm0, *I(0));    /* save I(0) */ \
 446     movdqa_r2m(xmm1, *I(1));    /* save I(1) */ \
 447     \
 448     movdqa_m2r(*I(6), xmm0);    /* load h3g3g3e3h2g2f2e2 */ \
 449     movdqa_r2r(xmm3, xmm1);     /* make a copy */ \
 450     \
 451     punpcklqdq_r2r(xmm0, xmm1); /* xmm1=h2g2f2e2d2c2b2a2 */ \
 452     punpckhqdq_r2r(xmm0, xmm3); /* xmm3=h3g3f3e3d3c3b3a3 */ \
 453     \
 454     movdqa_r2r(xmm2, xmm4);     /* make a copy */ \
 455     punpcklqdq_r2r(xmm5, xmm4); /* xmm4=h4g4f4e4d4c4b4a4 */ \
 456     \
 457     punpckhqdq_r2r(xmm5, xmm2); /* xmm2=h5g5f5e5d5c5b5a5 */ \
 458     movdqa_r2m(xmm1, *I(2));    /* save I(2) */ \
 459     \
 460     movdqa_r2m(xmm3, *I(3));    /* save I(3) */ \
 461     movdqa_r2m(xmm4, *I(4));    /* save I(4) */ \
 462     \
 463     movdqa_r2m(xmm2, *I(5));    /* save I(5) */ \
 464     movdqa_r2r(xmm7, xmm5);     /* make a copy */ \
 465     \
 466     punpcklqdq_r2r(xmm6, xmm5); /* xmm5=h6g6f6e6d6c6b6a6 */ \
 467     punpckhqdq_r2r(xmm6, xmm7); /* xmm7=h7g7f7e7d7c7b7a7 */ \
 468     \
 469     movdqa_r2m(xmm5, *I(6));    /* save I(6) */ \
 470     movdqa_r2m(xmm7, *I(7));    /* save I(7) */ \
 471     \
 472 } /* End of Transpose Macro */
 473
 474
 475 #define SSE2_Dequantize() {        \
 476     movdqa_m2r(*(eax), xmm0);     \
 477     \
 478     pmullw_m2r(*(ebx), xmm0);          /* xmm0 = 07 06 05 04 03 02 01 00 */ \
 479     movdqa_m2r(*(eax + 16), xmm1);     \
 480     \
 481     pmullw_m2r(*(ebx + 16), xmm1);     /* xmm1 = 17 16 15 14 13 12 11 10 */ \
 482     pshuflw_r2r(xmm0, xmm3, 0x078);    /* xmm3 = 07 06 05 04 01 03 02 00 */ \
 483     \
 484     movdqa_r2r(xmm1, xmm2);            /* xmm2 = 17 16 15 14 13 12 11 10 */ \
 485     movdqa_m2r(*(ecx), xmm7);          /* xmm7 = -- -- -- -- -- FF FF -- */ \
 486     \
 487     movdqa_m2r(*(eax + 32), xmm4);     \
 488     movdqa_m2r(*(eax + 64), xmm5);     \
 489     \
 490     pmullw_m2r(*(ebx + 32), xmm4);     /* xmm4 = 27 26 25 24 23 22 21 20 */ \
 491     pmullw_m2r(*(ebx + 64), xmm5);     /* xmm5 = 47 46 45 44 43 42 41 40 */ \
 492     \
 493     movdqa_m2r(*(ecx + 16), xmm6);     /* xmm6 = -- -- FF FF -- -- -- -- */ \
 494     pand_r2r(xmm2, xmm7);              /* xmm7 = -- -- -- -- -- 12 11 -- */ \
 495     \
 496     pand_r2r(xmm4, xmm6);              /* xmm6 = -- -- 25 24 -- -- -- -- */ \
 497     pxor_r2r(xmm7, xmm2);              /* xmm2 = 17 16 15 14 13 -- -- 10 */ \
 498     \
 499     pxor_r2r(xmm6, xmm4);              /* xmm4 = 27 26 -- -- 23 22 21 20 */ \
 500     pslldq_i2r(4, xmm7);               /* xmm7 = -- -- -- 12 11 -- -- -- */ \
 501     \
 502     pslldq_i2r(2, xmm6);               /* xmm6 = -- 25 24 -- -- -- -- -- */ \
 503     por_r2r(xmm6, xmm7);               /* xmm7 = -- 25 24 12 11 -- -- -- */ \
 504     \
 505     movdqa_m2r(*(ecx + 32), xmm0);     /* xmm0 = -- -- -- -- -- FF FF FF */ \
 506     movdqa_m2r(*(ecx + 48), xmm6);     /* xmm6 = -- -- -- -- FF -- -- -- */ \
 507     \
 508     pand_r2r(xmm3, xmm0);              /* xmm0 = -- -- -- -- -- 03 02 00 */ \
 509     pand_r2r(xmm5, xmm6);              /* xmm6 = -- -- -- -- 43 -- -- -- */ \
 510     \
 511     pxor_r2r(xmm0, xmm3);              /* xmm3 = 07 06 05 04 01 -- -- -- */ \
 512     pxor_r2r(xmm6, xmm5);              /* xmm5 = 47 46 45 44 -- 42 41 40 */ \
 513     \
 514     por_r2r(xmm7, xmm0);               /* xmm0 = -- 25 24 12 11 03 02 00 */ \
 515     pslldq_i2r(8, xmm6);               /* xmm6 = 43 -- -- -- -- -- -- -- */ \
 516     \
 517     por_r2r(xmm6, xmm0);               /* xmm0 = 43 25 24 12 11 03 02 00 */ \
 518     /* 02345 in use */ \
 519     \
 520     movdqa_m2r(*(ecx + 64 ), xmm1);    /* xmm1 = -- -- -- FF FF -- -- -- */ \
 521     pshuflw_r2r(xmm5, xmm5, 0x0B4);    /* xmm5 = 47 46 45 44 42 -- 41 40 */ \
 522     \
 523     movdqa_r2r(xmm1, xmm7);            /* xmm7 = -- -- -- FF FF -- -- -- */ \
 524     movdqa_r2r(xmm1, xmm6);            /* xmm6 = -- -- -- FF FF -- -- -- */ \
 525     \
 526     movdqa_r2m(xmm0, *(eax));          /* write  43 25 24 12 11 03 02 00 */ \
 527     pshufhw_r2r(xmm4, xmm4, 0x0C2);    /* xmm4 = 27 -- -- 26 23 22 21 20 */ \
 528     \
 529     pand_r2r(xmm4, xmm7);              /* xmm7 = -- -- -- 26 23 -- -- -- */ \
 530     pand_r2r(xmm5, xmm1);              /* xmm1 = -- -- -- 44 42 -- -- -- */ \
 531     \
 532     pxor_r2r(xmm7, xmm4);              /* xmm4 = 27 -- -- -- -- 22 21 20 */ \
 533     pxor_r2r(xmm1, xmm5);              /* xmm5 = 47 46 45 -- -- -- 41 40 */ \
 534     \
 535     pshuflw_r2r(xmm2, xmm2, 0x0C6);    /* xmm2 = 17 16 15 14 13 10 -- -- */ \
 536     movdqa_r2r(xmm6, xmm0);            /* xmm0 = -- -- -- FF FF -- -- -- */ \
 537     \
 538     pslldq_i2r(2, xmm7);               /* xmm7 = -- -- 26 23 -- -- -- -- */ \
 539     pslldq_i2r(6, xmm1);               /* xmm1 = 44 42 -- -- -- -- -- -- */ \
 540     \
 541     psrldq_i2r(2, xmm0);               /* xmm0 = -- -- -- -- FF FF -- -- */ \
 542     pand_r2r(xmm3, xmm6);              /* xmm6 = -- -- -- 04 01 -- -- -- */ \
 543     \
 544     pand_r2r(xmm2, xmm0);              /* xmm0 = -- -- -- -- 13 10 -- -- */ \
 545     pxor_r2r(xmm6, xmm3);              /* xmm3 = 07 06 05 -- -- -- -- -- */ \
 546     \
 547     pxor_r2r(xmm0, xmm2);              /* xmm2 = 17 16 15 14 -- -- -- -- */ \
 548     psrldq_i2r(6, xmm6);               /* xmm0 = -- -- -- -- -- -- 04 01 */ \
 549     \
 550     por_r2r(xmm7, xmm1);               /* xmm1 = 44 42 26 23 -- -- -- -- */ \
 551     por_r2r(xmm6, xmm0);               /* xmm1 = -- -- -- -- 13 10 04 01 */ \
 552     /* 12345 in use */    \
 553     por_r2r(xmm0, xmm1);               /* xmm1 = 44 42 26 23 13 10 04 01 */ \
 554     pshuflw_r2r(xmm4, xmm4, 0x093);    /* xmm4 = 27 -- -- -- 22 21 20 -- */ \
 555     \
 556     pshufhw_r2r(xmm4, xmm4, 0x093);    /* xmm4 = -- -- -- 27 22 21 20 -- */ \
 557     movdqa_r2m(xmm1, *(eax + 16));     /* write  44 42 26 23 13 10 04 01 */ \
 558     \
 559     pshufhw_r2r(xmm3, xmm3, 0x0D2);    /* xmm3 = 07 05 -- 06 -- -- -- -- */ \
 560     movdqa_m2r(*(ecx + 64), xmm0);     /* xmm0 = -- -- -- FF FF -- -- -- */ \
 561     \
 562     pand_r2r(xmm3, xmm0);              /* xmm0 = -- -- -- 06 -- -- -- -- */ \
 563     psrldq_i2r(12, xmm3);              /* xmm3 = -- -- -- -- -- -- 07 05 */ \
 564     \
 565     psrldq_i2r(8, xmm0);               /* xmm0 = -- -- -- -- -- -- -- 06 */ \
 566     \
 567     movdqa_m2r(*(ecx + 64), xmm6);     /* xmm6 = -- -- -- FF FF -- -- -- */ \
 568     movdqa_m2r(*(ecx + 96), xmm7);     /* xmm7 = -- -- -- -- FF FF -- -- */ \
 569     \
 570     pand_r2r(xmm4, xmm6);              /* xmm6 = -- -- -- 27 22 -- -- -- */ \
 571     pxor_r2r(xmm6, xmm4);              /* xmm4 = -- -- -- -- -- 21 20 -- */ \
 572     \
 573     por_r2r(xmm6, xmm3);               /* xmm3 = -- -- -- 27 22 -- 07 05 */ \
 574     pand_r2r(xmm4, xmm7);              /* xmm7 = -- -- -- -- -- 21 -- -- */ \
 575     \
 576     por_r2r(xmm7, xmm0);               /* xmm0 = -- -- -- -- -- 21 -- 06 */ \
 577     pxor_r2r(xmm7, xmm4);              /* xmm4 = -- -- -- -- -- -- 20 -- */ \
 578     \
 579     movdqa_m2r(*(ecx + 16 ), xmm6);    /* xmm6 = -- -- FF FF -- -- -- -- */ \
 580     movdqa_m2r(*(ecx + 64 ), xmm1);    /* xmm1 = -- -- -- FF FF -- -- -- */ \
 581     \
 582     pand_r2r(xmm2, xmm6);              /* xmm6 = -- -- 15 14 -- -- -- -- */ \
 583     pand_r2r(xmm6, xmm1);              /* xmm1 = -- -- -- 14 -- -- -- -- */ \
 584     \
 585     pxor_r2r(xmm6, xmm2);              /* xmm2 = 17 16 -- -- -- -- -- -- */ \
 586     pxor_r2r(xmm1, xmm6);              /* xmm6 = -- -- 15 -- -- -- -- -- */ \
 587     \
 588     psrldq_i2r(4, xmm1);               /* xmm1 = -- -- -- -- -- 14 -- -- */ \
 589     \
 590     psrldq_i2r(8, xmm6);               /* xmm6 = -- -- -- -- -- -- 15 -- */ \
 591     por_r2r(xmm1, xmm3);               /* xmm3 = -- -- -- 27 22 14 07 05 */ \
 592     \
 593     por_r2r(xmm6, xmm0);               /* xmm0 = -- -- -- -- -- 21 15 06 */ \
 594     pshufhw_r2r(xmm5, xmm5, 0x0E1);    /* xmm5 = 47 46 -- 45 -- -- 41 40 */ \
 595     \
 596     movdqa_m2r(*(ecx + 64), xmm1);     /* xmm1 = -- -- -- FF FF -- -- -- */ \
 597     pshuflw_r2r(xmm5, xmm5, 0x072);    /* xmm5 = 47 46 -- 45 41 -- 40 -- */ \
 598     \
 599     movdqa_r2r(xmm1, xmm6);            /* xmm6 = -- -- -- FF FF -- -- -- */ \
 600     pand_r2r(xmm5, xmm1);              /* xmm1 = -- -- -- 45 41 -- -- -- */ \
 601     \
 602     pxor_r2r(xmm1, xmm5);              /* xmm5 = 47 46 -- -- -- -- 40 -- */ \
 603     pslldq_i2r(4, xmm1);               /* xmm1 = -- 45 41 -- -- -- -- -- */ \
 604     \
 605     pshufd_r2r(xmm5, xmm5, 0x09C);     /* xmm5 = -- -- -- -- 47 46 40 -- */ \
 606     por_r2r(xmm1, xmm3);               /* xmm3 = -- 45 41 27 22 14 07 05 */ \
 607     \
 608     movdqa_m2r(*(eax + 96), xmm1);     /* xmm1 = 67 66 65 64 63 62 61 60 */ \
 609     pmullw_m2r(*(ebx + 96), xmm1);     \
 610     \
 611     movdqa_m2r(*(ecx), xmm7);          /* xmm7 = -- -- -- -- -- FF FF -- */ \
 612     \
 613     psrldq_i2r(8, xmm6);               /* xmm6 = -- -- -- -- -- -- -- FF */ \
 614     pand_r2r(xmm5, xmm7);              /* xmm7 = -- -- -- -- -- 46 40 -- */ \
 615     \
 616     pand_r2r(xmm1, xmm6);              /* xmm6 = -- -- -- -- -- -- -- 60 */ \
 617     pxor_r2r(xmm7, xmm5);              /* xmm5 = -- -- -- -- 47 -- -- -- */ \
 618     \
 619     pxor_r2r(xmm6, xmm1);              /* xmm1 = 67 66 65 64 63 62 61 -- */ \
 620     pslldq_i2r(2, xmm5);               /* xmm5 = -- -- -- 47 -- -- -- -- */ \
 621     \
 622     pslldq_i2r(14, xmm6);              /* xmm6 = 60 -- -- -- -- -- -- -- */ \
 623     por_r2r(xmm5, xmm4);               /* xmm4 = -- -- -- 47 -- -- 20 -- */ \
 624     \
 625     por_r2r(xmm6, xmm3);               /* xmm3 = 60 45 41 27 22 14 07 05 */ \
 626     pslldq_i2r(6, xmm7);               /* xmm7 = -- -- 46 40 -- -- -- -- */ \
 627     \
 628     movdqa_r2m(xmm3, *(eax+32));       /* write  60 45 41 27 22 14 07 05 */ \
 629     por_r2r(xmm7, xmm0);               /* xmm0 = -- -- 46 40 -- 21 15 06 */ \
 630     /* 0, 1, 2, 4 in use */    \
 631     movdqa_m2r(*(eax + 48), xmm3);     /* xmm3 = 37 36 35 34 33 32 31 30 */ \
 632     movdqa_m2r(*(eax + 80), xmm5);     /* xmm5 = 57 56 55 54 53 52 51 50 */ \
 633     \
 634     pmullw_m2r(*(ebx + 48), xmm3);     \
 635     pmullw_m2r(*(ebx + 80), xmm5);     \
 636     \
 637     movdqa_m2r(*(ecx + 64), xmm6);     /* xmm6 = -- -- -- FF FF -- -- -- */ \
 638     movdqa_m2r(*(ecx + 64), xmm7);     /* xmm7 = -- -- -- FF FF -- -- -- */ \
 639     \
 640     psrldq_i2r(8, xmm6);               /* xmm6 = -- -- -- -- -- -- -- FF */ \
 641     pslldq_i2r(8, xmm7);               /* xmm7 = FF -- -- -- -- -- -- -- */ \
 642     \
 643     pand_r2r(xmm3, xmm6);              /* xmm6 = -- -- -- -- -- -- -- 30 */ \
 644     pand_r2r(xmm5, xmm7);              /* xmm7 = 57 -- -- -- -- -- -- -- */ \
 645     \
 646     pxor_r2r(xmm6, xmm3);              /* xmm3 = 37 36 35 34 33 32 31 -- */ \
 647     pxor_r2r(xmm7, xmm5);              /* xmm5 = __ 56 55 54 53 52 51 50 */ \
 648     \
 649     pslldq_i2r(6, xmm6);               /* xmm6 = -- -- -- -- 30 -- -- -- */ \
 650     psrldq_i2r(2, xmm7);               /* xmm7 = -- 57 -- -- -- -- -- -- */ \
 651     \
 652     por_r2r(xmm7, xmm6);               /* xmm6 = -- 57 -- -- 30 -- -- -- */ \
 653     movdqa_m2r(*(ecx), xmm7);          /* xmm7 = -- -- -- -- -- FF FF -- */ \
 654     \
 655     por_r2r(xmm6, xmm0);               /* xmm0 = -- 57 46 40 30 21 15 06 */ \
 656     psrldq_i2r(2, xmm7);               /* xmm7 = -- -- -- -- -- -- FF FF */ \
 657     \
 658     movdqa_r2r(xmm2, xmm6);            /* xmm6 = 17 16 -- -- -- -- -- -- */ \
 659     pand_r2r(xmm1, xmm7);              /* xmm7 = -- -- -- -- -- -- 61 -- */ \
 660     \
 661     pslldq_i2r(2, xmm6);               /* xmm6 = 16 -- -- -- -- -- -- -- */ \
 662     psrldq_i2r(14, xmm2);              /* xmm2 = -- -- -- -- -- -- -- 17 */ \
 663     \
 664     pxor_r2r(xmm7, xmm1);              /* xmm1 = 67 66 65 64 63 62 -- -- */ \
 665     pslldq_i2r(12, xmm7);              /* xmm7 = 61 -- -- -- -- -- -- -- */ \
 666     \
 667     psrldq_i2r(14, xmm6);              /* xmm6 = -- -- -- -- -- -- -- 16 */ \
 668     por_r2r(xmm6, xmm4);               /* xmm4 = -- -- -- 47 -- -- 20 16 */ \
 669     \
 670     por_r2r(xmm7, xmm0);               /* xmm0 = 61 57 46 40 30 21 15 06 */ \
 671     movdqa_m2r(*(ecx), xmm6);          /* xmm6 = -- -- -- -- -- FF FF -- */ \
 672     \
 673     psrldq_i2r(2, xmm6);               /* xmm6 = -- -- -- -- -- -- FF FF */ \
 674     movdqa_r2m(xmm0, *(eax+48));       /* write  61 57 46 40 30 21 15 06 */ \
 675     /* 1, 2, 3, 4, 5 in use */\
 676     movdqa_m2r(*(ecx), xmm0);          /* xmm0 = -- -- -- -- -- FF FF -- */ \
 677     pand_r2r(xmm3, xmm6);              /* xmm6 = -- -- -- -- -- -- 31 -- */ \
 678     \
 679     movdqa_r2r(xmm3, xmm7);            /* xmm7 = 37 36 35 34 33 32 31 -- */ \
 680     pxor_r2r(xmm6, xmm3);              /* xmm3 = 37 36 35 34 33 32 -- -- */ \
 681     \
 682     pslldq_i2r(2, xmm3);               /* xmm3 = 36 35 34 33 32 -- -- -- */ \
 683     pand_r2r(xmm1, xmm0);              /* xmm0 = -- -- -- -- -- 62 -- -- */ \
 684     \
 685     psrldq_i2r(14, xmm7);              /* xmm7 = -- -- -- -- -- -- -- 37 */ \
 686     pxor_r2r(xmm0, xmm1);              /* xmm1 = 67 66 65 64 63 -- -- -- */ \
 687     \
 688     por_r2r(xmm7, xmm6);               /* xmm6 = -- -- -- -- -- -- 31 37 */ \
 689     movdqa_m2r(*(ecx + 64), xmm7);     /* xmm7 = -- -- -- FF FF -- -- -- */ \
 690     \
 691     pshuflw_r2r(xmm6, xmm6, 0x01E);    /* xmm6 = -- -- -- -- 37 31 -- -- */ \
 692     pslldq_i2r(6, xmm7);               /* xmm7 = FF FF -- -- -- -- -- -- */ \
 693     \
 694     por_r2r(xmm6, xmm4);               /* xmm4 = -- -- -- 47 37 31 20 16 */ \
 695     pand_r2r(xmm5, xmm7);              /* xmm7 = -- 56 -- -- -- -- -- -- */ \
 696     \
 697     pslldq_i2r(8, xmm0);               /* xmm0 = -- 62 -- -- -- -- -- -- */ \
 698     pxor_r2r(xmm7, xmm5);              /* xmm5 = -- -- 55 54 53 52 51 50 */ \
 699     \
 700     psrldq_i2r(2, xmm7);               /* xmm7 = -- -- 56 -- -- -- -- -- */ \
 701     \
 702     pshufhw_r2r(xmm3, xmm3, 0x087);    /* xmm3 = 35 33 34 36 32 -- -- -- */ \
 703     por_r2r(xmm7, xmm0);               /* xmm0 = -- 62 56 -- -- -- -- -- */ \
 704     \
 705     movdqa_m2r(*(eax + 112), xmm7);    /* xmm7 = 77 76 75 74 73 72 71 70 */ \
 706     pmullw_m2r(*(ebx + 112), xmm7);     \
 707     \
 708     movdqa_m2r(*(ecx + 64), xmm6);     /* xmm6 = -- -- -- FF FF -- -- -- */ \
 709     por_r2r(xmm0, xmm4);               /* xmm4 = -- 62 56 47 37 31 20 16 */ \
 710     \
 711     pshuflw_r2r(xmm7, xmm7, 0x0E1);    /* xmm7 = 77 76 75 74 73 72 70 71 */ \
 712     psrldq_i2r(8, xmm6);               /* xmm6 = -- -- -- -- -- -- -- FF */ \
 713     \
 714     movdqa_m2r(*(ecx + 64), xmm0);     /* xmm0 = -- -- -- FF FF -- -- -- */ \
 715     pand_r2r(xmm7, xmm6);              /* xmm6 = -- -- -- -- -- -- -- 71 */ \
 716     \
 717     pand_r2r(xmm3, xmm0);              /* xmm0 = -- -- -- 36 32 -- -- -- */ \
 718     pxor_r2r(xmm6, xmm7);              /* xmm7 = 77 76 75 74 73 72 70 -- */ \
 719     \
 720     pxor_r2r(xmm0, xmm3);              /* xmm3 = 35 33 34 -- -- -- -- -- */ \
 721     pslldq_i2r(14, xmm6);              /* xmm6 = 71 -- -- -- -- -- -- -- */ \
 722     \
 723     psrldq_i2r(4, xmm0);               /* xmm0 = -- -- -- -- -- 36 32 -- */ \
 724     por_r2r(xmm6, xmm4);               /* xmm4 = 71 62 56 47 37 31 20 16 */ \
 725     \
 726     por_r2r(xmm0, xmm2);               /* xmm2 = -- -- -- -- -- 36 32 17 */ \
 727     movdqa_r2m(xmm4, *(eax + 64));     /* write  71 62 56 47 37 31 20 16 */ \
 728     /* 1, 2, 3, 5, 7 in use */ \
 729     movdqa_m2r(*(ecx + 80), xmm6);     /* xmm6 = -- -- FF -- -- -- -- FF */ \
 730     pshufhw_r2r(xmm7, xmm7, 0x0D2);    /* xmm7 = 77 75 74 76 73 72 70 __ */ \
 731     \
 732     movdqa_m2r(*(ecx), xmm4);          /* xmm4 = -- -- -- -- -- FF FF -- */ \
 733     movdqa_m2r(*(ecx+48), xmm0);       /* xmm0 = -- -- -- -- FF -- -- -- */ \
 734     \
 735     pand_r2r(xmm5, xmm6);              /* xmm6 = -- -- 55 -- -- -- -- 50 */ \
 736     pand_r2r(xmm7, xmm4);              /* xmm4 = -- -- -- -- -- 72 70 -- */ \
 737     \
 738     pand_r2r(xmm1, xmm0);              /* xmm0 = -- -- -- -- 63 -- -- -- */ \
 739     pxor_r2r(xmm6, xmm5);              /* xmm5 = -- -- -- 54 53 52 51 -- */ \
 740     \
 741     pxor_r2r(xmm4, xmm7);              /* xmm7 = 77 75 74 76 73 -- -- -- */ \
 742     pxor_r2r(xmm0, xmm1);              /* xmm1 = 67 66 65 64 -- -- -- -- */ \
 743     \
 744     pshuflw_r2r(xmm6, xmm6, 0x02B);    /* xmm6 = -- -- 55 -- 50 -- -- -- */ \
 745     pslldq_i2r(10, xmm4);              /* xmm4 = 72 20 -- -- -- -- -- -- */ \
 746     \
 747     pshufhw_r2r(xmm6, xmm6, 0x0B1);    /* xmm6 = -- -- -- 55 50 -- -- -- */ \
 748     pslldq_i2r(4, xmm0);               /* xmm0 = -- -- 63 -- -- -- -- -- */ \
 749     \
 750     por_r2r(xmm4, xmm6);               /* xmm6 = 72 70 -- 55 50 -- -- -- */ \
 751     por_r2r(xmm0, xmm2);               /* xmm2 = -- -- 63 -- -- 36 32 17 */ \
 752     \
 753     por_r2r(xmm6, xmm2);               /* xmm2 = 72 70 64 55 50 36 32 17 */ \
 754     pshufhw_r2r(xmm1, xmm1, 0x0C9);    /* xmm1 = 67 64 66 65 -- -- -- -- */ \
 755     \
 756     movdqa_r2r(xmm3, xmm6);            /* xmm6 = 35 33 34 -- -- -- -- -- */ \
 757     movdqa_r2m(xmm2, *(eax+80));       /* write  72 70 64 55 50 36 32 17 */ \
 758     \
 759     psrldq_i2r(12, xmm6);              /* xmm6 = -- -- -- -- -- -- 35 33 */ \
 760     pslldq_i2r(4, xmm3);               /* xmm3 = 34 -- -- -- -- -- -- -- */ \
 761     \
 762     pshuflw_r2r(xmm5, xmm5, 0x04E);    /* xmm5 = -- -- -- 54 51 -- 53 52 */ \
 763     movdqa_r2r(xmm7, xmm4);            /* xmm4 = 77 75 74 76 73 -- -- -- */ \
 764     \
 765     movdqa_r2r(xmm5, xmm2);            /* xmm2 = -- -- -- 54 51 -- 53 52 */ \
 766     psrldq_i2r(10, xmm7);              /* xmm7 = -- -- -- -- -- 77 75 74 */ \
 767     \
 768     pslldq_i2r(6, xmm4);               /* xmm4 = 76 73 -- -- -- -- -- -- */ \
 769     pslldq_i2r(12, xmm2);              /* xmm2 = 53 52 -- -- -- -- -- -- */ \
 770     \
 771     movdqa_r2r(xmm1, xmm0);            /* xmm0 = 67 64 66 65 -- -- -- -- */ \
 772     psrldq_i2r(12, xmm1);              /* xmm1 = -- -- -- -- -- -- 67 64 */ \
 773     \
 774     psrldq_i2r(6, xmm5);               /* xmm5 = -- -- -- -- -- -- 54 51 */ \
 775     psrldq_i2r(14, xmm3);              /* xmm3 = -- -- -- -- -- -- -- 34 */ \
 776     \
 777     pslldq_i2r(10, xmm7);              /* xmm7 = 77 75 74 -- -- -- -- -- */ \
 778     por_r2r(xmm6, xmm4);               /* xmm4 = 76 73 -- -- -- -- 35 33 */ \
 779     \
 780     psrldq_i2r(10, xmm2);              /* xmm2 = -- -- -- -- -- 53 52 -- */ \
 781     pslldq_i2r(4, xmm0);               /* xmm0 = 66 65 -- -- -- -- -- -- */ \
 782     \
 783     pslldq_i2r(8, xmm1);               /* xmm1 = -- -- 67 64 -- -- -- -- */ \
 784     por_r2r(xmm7, xmm3);               /* xmm3 = 77 75 74 -- -- -- -- 34 */ \
 785     \
 786     psrldq_i2r(6, xmm0);               /* xmm0 = -- -- -- 66 65 -- -- -- */ \
 787     pslldq_i2r(4, xmm5);               /* xmm5 = -- -- -- -- 54 51 -- -- */ \
 788     \
 789     por_r2r(xmm1, xmm4);               /* xmm4 = 76 73 67 64 -- -- 35 33 */ \
 790     por_r2r(xmm2, xmm3);               /* xmm3 = 77 75 74 -- -- 53 52 34 */ \
 791     \
 792     por_r2r(xmm5, xmm4);               /* xmm4 = 76 73 67 64 54 51 35 33 */ \
 793     por_r2r(xmm0, xmm3);               /* xmm3 = 77 75 74 66 65 53 52 34 */ \
 794     \
 795     movdqa_r2m(xmm4, *(eax+96));       /* write  76 73 67 64 54 51 35 33 */ \
 796     movdqa_r2m(xmm3, *(eax+112));      /* write  77 75 74 66 65 53 52 34 */ \
 797     \
 798 } /* end of SSE2_Dequantize Macro */
 799
 800
 801 void ff_vp3_idct_sse2(int16_t *input_data)
 802 {
 803     unsigned char *input_bytes = (unsigned char *)input_data;
 804     unsigned char *output_data_bytes = (unsigned char *)input_data;
 805     const unsigned char *idct_data_bytes = (const unsigned char *)SSE2_idct_data;
 806     const unsigned char *Eight = (const unsigned char *)eight_data;
 807
 808 #define eax input_bytes
 809 //#define ebx dequant_matrix_bytes
 810 #define ecx dequant_const_bytes
 811 #define edx idct_data_bytes
 812
 813 #define I(i) (eax + 16 * i)
 814 #define O(i) (ebx + 16 * i)
 815 #define C(i) (edx + 16 * (i-1))
 816
 817  //   SSE2_Dequantize();
 818
 819 #undef ebx
 820 #define ebx output_data_bytes
 821
 822     SSE2_Row_IDCT();
 823
 824     SSE2_Transpose();
 825
 826     SSE2_Column_IDCT();
 827 }
 828
 829 void ff_vp3_idct_put_sse2(uint8_t *dest, int line_size, DCTELEM *block)
 830 {
 831     ff_vp3_idct_sse2(block);
 832     put_signed_pixels_clamped_mmx(block, dest, line_size);
 833 }
 834
 835 void ff_vp3_idct_add_sse2(uint8_t *dest, int line_size, DCTELEM *block)
 836 {
 837     ff_vp3_idct_sse2(block);
 838     add_pixels_clamped_mmx(block, dest, line_size);
 839 }