vp3dsp_sse2.c 35.6 KB
Newer Older
1 2 3
/*
 * Copyright (C) 2004 the ffmpeg project
 *
4 5 6
 * This file is part of FFmpeg.
 *
 * FFmpeg is free software; you can redistribute it and/or
7 8
 * modify it under the terms of the GNU Lesser General Public
 * 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 13 14 15 16
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * 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 23 24 25
 */

/**
 * @file vp3dsp_sse2.c
 * SSE2-optimized functions cribbed from the original VP3 source code.
 */

26
#include "libavcodec/dsputil.h"
27 28
#include "mmx.h"

29
static DECLARE_ALIGNED_16(const unsigned short, SSE2_dequant_const[]) =
30 31 32 33 34 35 36 37 38 39
{
    0,65535,65535,0,0,0,0,0,    // 0x0000 0000 0000 0000 0000 FFFF FFFF 0000
    0,0,0,0,65535,65535,0,0,    // 0x0000 0000 FFFF FFFF 0000 0000 0000 0000
    65535,65535,65535,0,0,0,0,0,// 0x0000 0000 0000 0000 0000 FFFF FFFF FFFF
    0,0,0,65535,0,0,0,0,        // 0x0000 0000 0000 0000 FFFF 0000 0000 0000
    0,0,0,65535,65535,0,0,0,    // 0x0000 0000 0000 FFFF FFFF 0000 0000 0000
    65535,0,0,0,0,65535,0,0,    // 0x0000 0000 FFFF 0000 0000 0000 0000 FFFF
    0,0,65535,65535, 0,0,0,0    // 0x0000 0000 0000 0000 FFFF FFFF 0000 0000
};

40
static DECLARE_ALIGNED_16(const unsigned int, eight_data[]) =
41 42 43
{
    0x00080008,
    0x00080008,
44
    0x00080008,
45 46
    0x00080008
};
47

48
static DECLARE_ALIGNED_16(const unsigned short, SSE2_idct_data[7 * 8]) =
49
{
50 51 52 53 54 55
    64277,64277,64277,64277,64277,64277,64277,64277,
    60547,60547,60547,60547,60547,60547,60547,60547,
    54491,54491,54491,54491,54491,54491,54491,54491,
    46341,46341,46341,46341,46341,46341,46341,46341,
    36410,36410,36410,36410,36410,36410,36410,36410,
    25080,25080,25080,25080,25080,25080,25080,25080,
56 57 58 59 60 61
    12785,12785,12785,12785,12785,12785,12785,12785
};


#define SSE2_Column_IDCT() {        \
    \
62 63
    movdqa_m2r(*I(3), xmm2);     /* xmm2 = i3 */ \
    movdqa_m2r(*C(3), xmm6);     /* xmm6 = c3 */ \
64
    \
65 66
    movdqa_r2r(xmm2, xmm4);      /* xmm4 = i3 */ \
    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 81 82 83 84 85
    \
    /* all registers are in use */ \
    \
    paddw_r2r(xmm2, xmm4);       /* xmm4 = c3 * i3 */ \
    paddw_r2r(xmm7, xmm6);       /* xmm6 = c3 * i5 */ \
    \
    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 92 93 94
    \
    pmulhw_r2r(xmm3, xmm0);      /* xmm0 = c1 * i1 - i1 */ \
    paddsw_r2r(xmm7, xmm4);      /* xmm4 = c3 * i3 + c5 * i5 = C */ \
    \
    pmulhw_r2r(xmm1, xmm5);      /* xmm5 = c1 * i7 - i7 */ \
95
    movdqa_m2r(*C(7), xmm7);     /* xmm7 = c7 */ \
96 97 98 99 100
    \
    psubsw_r2r(xmm2, xmm6);      /* xmm6 = c3 * i5 - c5 * i3 = D */ \
    paddw_r2r(xmm3, xmm0);       /* xmm0 = c1 * i1 */ \
    \
    pmulhw_r2r(xmm7, xmm3);      /* xmm3 = c7 * i1 */ \
101
    movdqa_m2r(*I(2), xmm2);     /* xmm2 = i2 */ \
102 103 104 105
    \
    pmulhw_r2r(xmm1, xmm7);      /* xmm7 = c7 * i7 */ \
    paddw_r2r(xmm1, xmm5);       /* xmm5 = c1 * i7 */ \
    \
106
    movdqa_r2r(xmm2, xmm1);      /* xmm1 = i2 */ \
107 108 109
    pmulhw_m2r(*C(2), xmm2);     /* xmm2 = i2 * c2 -i2 */ \
    \
    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 */ \
114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130
    \
    psubsw_r2r(xmm4, xmm0);      /* xmm0 = A - C */ \
    pmulhw_m2r(*C(2), xmm5);     /* xmm5 = c2 * i6 - i6 */ \
    \
    paddw_r2r(xmm1, xmm2);       /* xmm2 = i2 * c2 */ \
    pmulhw_m2r(*C(6), xmm1);     /* xmm1 = c6 * i2 */ \
    \
    paddsw_r2r(xmm4, xmm4);      /* xmm4 = C + C */ \
    paddsw_r2r(xmm0, xmm4);      /* xmm4 = A + C = C. */ \
    \
    psubsw_r2r(xmm6, xmm3);      /* xmm3 = B - D */ \
    paddw_r2r(xmm7, xmm5);       /* xmm5 = c2 * i6 */ \
    \
    paddsw_r2r(xmm6, xmm6);      /* xmm6 = D + D */ \
    pmulhw_m2r(*C(6), xmm7);     /* xmm7 = c6 * i6 */ \
    \
    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 */ \
137 138 139
    pmulhw_r2r(xmm4, xmm3);      /* xmm3 = ( c4 -1 ) * ( B - D ) */ \
    \
    paddsw_r2r(xmm2, xmm7);      /* xmm7 = c2 * i2 + c6 * i6 = G */ \
140
    movdqa_r2m(xmm6, *I(2));     /* Save D. at I(2) */ \
141
    \
142 143
    movdqa_r2r(xmm0, xmm2);      /* xmm2 = A - C */ \
    movdqa_m2r(*I(0), xmm6);     /* xmm6 = i0 */ \
144 145 146 147
    \
    pmulhw_r2r(xmm4, xmm0);      /* xmm0 = ( c4 - 1 ) * ( A - C ) = A. */ \
    paddw_r2r(xmm3, xmm5);       /* xmm5 = c4 * ( B - D ) = B. */ \
    \
148
    movdqa_m2r(*I(4), xmm3);     /* xmm3 = i4 */ \
149 150 151 152 153
    psubsw_r2r(xmm1, xmm5);      /* xmm5 = B. - H = B.. */ \
    \
    paddw_r2r(xmm0, xmm2);       /* xmm2 = c4 * ( A - C) = A. */ \
    psubsw_r2r(xmm3, xmm6);      /* xmm6 = i0 - i4 */ \
    \
154
    movdqa_r2r(xmm6, xmm0);      /* xmm0 = i0 - i4 */ \
155 156 157 158 159 160 161 162 163 164 165 166 167 168
    pmulhw_r2r(xmm4, xmm6);      /* xmm6 = (c4 - 1) * (i0 - i4) = F */ \
    \
    paddsw_r2r(xmm3, xmm3);      /* xmm3 = i4 + i4 */ \
    paddsw_r2r(xmm1, xmm1);      /* xmm1 = H + H */ \
    \
    paddsw_r2r(xmm0, xmm3);      /* xmm3 = i0 + i4 */ \
    paddsw_r2r(xmm5, xmm1);      /* xmm1 = B. + H = H. */ \
    \
    pmulhw_r2r(xmm3, xmm4);      /* xmm4 = ( c4 - 1 ) * ( i0 + i4 )  */ \
    paddw_r2r(xmm0, xmm6);       /* xmm6 = c4 * ( i0 - i4 ) */ \
    \
    psubsw_r2r(xmm2, xmm6);      /* xmm6 = F - A. = F. */ \
    paddsw_r2r(xmm2, xmm2);      /* xmm2 = A. + A. */ \
    \
169
    movdqa_m2r(*I(1), xmm0);     /* Load        C. from I(1) */ \
170 171 172 173 174 175 176 177 178 179 180 181 182 183
    paddsw_r2r(xmm6, xmm2);      /* xmm2 = F + A. = A.. */ \
    \
    paddw_r2r(xmm3, xmm4);       /* xmm4 = c4 * ( i0 + i4 ) = 3 */ \
    psubsw_r2r(xmm1, xmm2);      /* xmm2 = A.. - H. = R2 */ \
    \
    paddsw_m2r(*Eight, xmm2);    /* Adjust R2 and R1 before shifting */ \
    paddsw_r2r(xmm1, xmm1);      /* xmm1 = H. + H. */ \
    \
    paddsw_r2r(xmm2, xmm1);      /* xmm1 = A.. + H. = R1 */ \
    psraw_i2r(4, xmm2);          /* xmm2 = op2 */ \
    \
    psubsw_r2r(xmm7, xmm4);      /* xmm4 = E - G = E. */ \
    psraw_i2r(4, xmm1);          /* xmm1 = op1 */ \
    \
184
    movdqa_m2r(*I(2), xmm3);     /* Load D. from I(2) */ \
185 186
    paddsw_r2r(xmm7, xmm7);      /* xmm7 = G + G */ \
    \
187
    movdqa_r2m(xmm2, *O(2));     /* Write out op2 */ \
188 189
    paddsw_r2r(xmm4, xmm7);      /* xmm7 = E + G = G. */ \
    \
190
    movdqa_r2m(xmm1, *O(1));     /* Write out op1 */ \
191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207
    psubsw_r2r(xmm3, xmm4);      /* xmm4 = E. - D. = R4 */ \
    \
    paddsw_m2r(*Eight, xmm4);    /* Adjust R4 and R3 before shifting */ \
    paddsw_r2r(xmm3, xmm3);      /* xmm3 = D. + D. */ \
    \
    paddsw_r2r(xmm4, xmm3);      /* xmm3 = E. + D. = R3 */ \
    psraw_i2r(4, xmm4);          /* xmm4 = op4 */ \
    \
    psubsw_r2r(xmm5, xmm6);      /* xmm6 = F. - B..= R6 */ \
    psraw_i2r(4, xmm3);          /* xmm3 = op3 */ \
    \
    paddsw_m2r(*Eight, xmm6);    /* Adjust R6 and R5 before shifting */ \
    paddsw_r2r(xmm5, xmm5);      /* xmm5 = B.. + B.. */ \
    \
    paddsw_r2r(xmm6, xmm5);      /* xmm5 = F. + B.. = R5 */ \
    psraw_i2r(4, xmm6);          /* xmm6 = op6 */ \
    \
208
    movdqa_r2m(xmm4, *O(4));     /* Write out op4 */ \
209 210
    psraw_i2r(4, xmm5);          /* xmm5 = op5 */ \
    \
211
    movdqa_r2m(xmm3, *O(3));     /* Write out op3 */ \
212 213 214 215 216 217 218 219
    psubsw_r2r(xmm0, xmm7);      /* xmm7 = G. - C. = R7 */ \
    \
    paddsw_m2r(*Eight, xmm7);    /* Adjust R7 and R0 before shifting */ \
    paddsw_r2r(xmm0, xmm0);      /* xmm0 = C. + C. */ \
    \
    paddsw_r2r(xmm7, xmm0);      /* xmm0 = G. + C. */ \
    psraw_i2r(4, xmm7);          /* xmm7 = op7 */ \
    \
220
    movdqa_r2m(xmm6, *O(6));     /* Write out op6 */ \
221 222
    psraw_i2r(4, xmm0);          /* xmm0 = op0 */ \
    \
223 224
    movdqa_r2m(xmm5, *O(5));     /* Write out op5 */ \
    movdqa_r2m(xmm7, *O(7));     /* Write out op7 */ \
225
    \
226
    movdqa_r2m(xmm0, *O(0));     /* Write out op0 */ \
227 228 229 230 231 232
    \
} /* End of SSE2_Column_IDCT macro */


#define SSE2_Row_IDCT() {        \
    \
233 234
    movdqa_m2r(*I(3), xmm2);     /* xmm2 = i3 */ \
    movdqa_m2r(*C(3), xmm6);     /* xmm6 = c3 */ \
235
    \
236 237
    movdqa_r2r(xmm2, xmm4);      /* xmm4 = i3 */ \
    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 */ \
250 251 252 253 254 255 256
    \
    /* all registers are in use */ \
    \
    paddw_r2r(xmm2, xmm4);      /* xmm4 = c3 * i3 */ \
    paddw_r2r(xmm7, xmm6);      /* xmm6 = c3 * i5 */ \
    \
    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 */ \
261 262 263 264 265
    \
    pmulhw_r2r(xmm3, xmm0);     /* xmm0 = c1 * i1 - i1 */ \
    paddsw_r2r(xmm7, xmm4);     /* xmm4 = c3 * i3 + c5 * i5 = C */ \
    \
    pmulhw_r2r(xmm1, xmm5);     /* xmm5 = c1 * i7 - i7 */ \
266
    movdqa_m2r(*C(7), xmm7);    /* xmm7 = c7 */ \
267 268 269 270 271
    \
    psubsw_r2r(xmm2, xmm6);     /* xmm6 = c3 * i5 - c5 * i3 = D */ \
    paddw_r2r(xmm3, xmm0);      /* xmm0 = c1 * i1 */ \
    \
    pmulhw_r2r(xmm7, xmm3);     /* xmm3 = c7 * i1 */ \
272
    movdqa_m2r(*I(2), xmm2);    /* xmm2 = i2 */ \
273 274 275 276
    \
    pmulhw_r2r(xmm1, xmm7);     /* xmm7 = c7 * i7 */ \
    paddw_r2r(xmm1, xmm5);      /* xmm5 = c1 * i7 */ \
    \
277
    movdqa_r2r(xmm2, xmm1);     /* xmm1 = i2 */ \
278 279 280
    pmulhw_m2r(*C(2), xmm2);    /* xmm2 = i2 * c2 -i2 */ \
    \
    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 */ \
285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301
    \
    psubsw_r2r(xmm4, xmm0);     /* xmm0 = A - C */ \
    pmulhw_m2r(*C(2), xmm5);    /* xmm5 = c2 * i6 - i6 */ \
    \
    paddw_r2r(xmm1, xmm2);      /* xmm2 = i2 * c2 */ \
    pmulhw_m2r(*C(6), xmm1);    /* xmm1 = c6 * i2 */ \
    \
    paddsw_r2r(xmm4, xmm4);     /* xmm4 = C + C */ \
    paddsw_r2r(xmm0, xmm4);     /* xmm4 = A + C = C. */ \
    \
    psubsw_r2r(xmm6, xmm3);     /* xmm3 = B - D */ \
    paddw_r2r(xmm7, xmm5);      /* xmm5 = c2 * i6 */ \
    \
    paddsw_r2r(xmm6, xmm6);     /* xmm6 = D + D */ \
    pmulhw_m2r(*C(6), xmm7);    /* xmm7 = c6 * i6 */ \
    \
    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 */ \
308 309 310
    pmulhw_r2r(xmm4, xmm3);     /* xmm3 = ( c4 -1 ) * ( B - D ) */ \
    \
    paddsw_r2r(xmm2, xmm7);     /* xmm7 = c2 * i2 + c6 * i6 = G */ \
311
    movdqa_r2m(xmm6, *I(2));    /* Save D. at I(2) */ \
312
    \
313 314
    movdqa_r2r(xmm0, xmm2);     /* xmm2 = A - C */ \
    movdqa_m2r(*I(0), xmm6);    /* xmm6 = i0 */ \
315 316 317 318
    \
    pmulhw_r2r(xmm4, xmm0);     /* xmm0 = ( c4 - 1 ) * ( A - C ) = A. */ \
    paddw_r2r(xmm3, xmm5);      /* xmm5 = c4 * ( B - D ) = B. */ \
    \
319
    movdqa_m2r(*I(4), xmm3);    /* xmm3 = i4 */ \
320 321 322 323 324
    psubsw_r2r(xmm1, xmm5);     /* xmm5 = B. - H = B.. */ \
    \
    paddw_r2r(xmm0, xmm2);      /* xmm2 = c4 * ( A - C) = A. */ \
    psubsw_r2r(xmm3, xmm6);     /* xmm6 = i0 - i4 */ \
    \
325
    movdqa_r2r(xmm6, xmm0);     /* xmm0 = i0 - i4 */ \
326 327 328 329 330 331 332 333 334 335 336 337 338 339
    pmulhw_r2r(xmm4, xmm6);     /* xmm6 = ( c4 - 1 ) * ( i0 - i4 ) = F */ \
    \
    paddsw_r2r(xmm3, xmm3);     /* xmm3 = i4 + i4 */ \
    paddsw_r2r(xmm1, xmm1);     /* xmm1 = H + H */ \
    \
    paddsw_r2r(xmm0, xmm3);     /* xmm3 = i0 + i4 */ \
    paddsw_r2r(xmm5, xmm1);     /* xmm1 = B. + H = H. */ \
    \
    pmulhw_r2r(xmm3, xmm4);     /* xmm4 = ( c4 - 1 ) * ( i0 + i4 )  */ \
    paddw_r2r(xmm0, xmm6);      /* xmm6 = c4 * ( i0 - i4 ) */ \
    \
    psubsw_r2r(xmm2, xmm6);     /* xmm6 = F - A. = F. */ \
    paddsw_r2r(xmm2, xmm2);     /* xmm2 = A. + A. */ \
    \
340
    movdqa_m2r(*I(1), xmm0);    /* Load C. from I(1) */ \
341 342 343 344 345 346 347 348 349 350
    paddsw_r2r(xmm6, xmm2);     /* xmm2 = F + A. = A.. */ \
    \
    paddw_r2r(xmm3, xmm4);      /* xmm4 = c4 * ( i0 + i4 ) = 3 */ \
    psubsw_r2r(xmm1, xmm2);     /* xmm2 = A.. - H. = R2 */ \
    \
    paddsw_r2r(xmm1, xmm1);     /* xmm1 = H. + H. */ \
    paddsw_r2r(xmm2, xmm1);     /* xmm1 = A.. + H. = R1 */ \
    \
    psubsw_r2r(xmm7, xmm4);     /* xmm4 = E - G = E. */ \
    \
351
    movdqa_m2r(*I(2), xmm3);    /* Load D. from I(2) */ \
352 353
    paddsw_r2r(xmm7, xmm7);     /* xmm7 = G + G */ \
    \
354
    movdqa_r2m(xmm2, *I(2));    /* Write out op2 */ \
355 356
    paddsw_r2r(xmm4, xmm7);     /* xmm7 = E + G = G. */ \
    \
357
    movdqa_r2m(xmm1, *I(1));    /* Write out op1 */ \
358 359 360 361 362 363 364 365 366 367 368 369
    psubsw_r2r(xmm3, xmm4);     /* xmm4 = E. - D. = R4 */ \
    \
    paddsw_r2r(xmm3, xmm3);     /* xmm3 = D. + D. */ \
    \
    paddsw_r2r(xmm4, xmm3);     /* xmm3 = E. + D. = R3 */ \
    \
    psubsw_r2r(xmm5, xmm6);     /* xmm6 = F. - B..= R6 */ \
    \
    paddsw_r2r(xmm5, xmm5);     /* xmm5 = B.. + B.. */ \
    \
    paddsw_r2r(xmm6, xmm5);     /* xmm5 = F. + B.. = R5 */ \
    \
370
    movdqa_r2m(xmm4, *I(4));    /* Write out op4 */ \
371
    \
372
    movdqa_r2m(xmm3, *I(3));    /* Write out op3 */ \
373 374 375 376 377 378
    psubsw_r2r(xmm0, xmm7);     /* xmm7 = G. - C. = R7 */ \
    \
    paddsw_r2r(xmm0, xmm0);     /* xmm0 = C. + C. */ \
    \
    paddsw_r2r(xmm7, xmm0);     /* xmm0 = G. + C. */ \
    \
379
    movdqa_r2m(xmm6, *I(6));    /* Write out op6 */ \
380
    \
381 382
    movdqa_r2m(xmm5, *I(5));    /* Write out op5 */ \
    movdqa_r2m(xmm7, *I(7));    /* Write out op7 */ \
383
    \
384
    movdqa_r2m(xmm0, *I(0));    /* Write out op0 */ \
385 386 387 388 389 390
    \
} /* End of SSE2_Row_IDCT macro */


#define SSE2_Transpose() {    \
    \
391 392
    movdqa_m2r(*I(4), xmm4);    /* xmm4=e7e6e5e4e3e2e1e0 */ \
    movdqa_m2r(*I(5), xmm0);    /* xmm4=f7f6f5f4f3f2f1f0 */ \
393
    \
394
    movdqa_r2r(xmm4, xmm5);     /* make a copy */ \
395 396 397
    punpcklwd_r2r(xmm0, xmm4);  /* xmm4=f3e3f2e2f1e1f0e0 */ \
    \
    punpckhwd_r2r(xmm0, xmm5);  /* xmm5=f7e7f6e6f5e5f4e4 */ \
398
    movdqa_m2r(*I(6), xmm6);    /* xmm6=g7g6g5g4g3g2g1g0 */ \
399
    \
400 401
    movdqa_m2r(*I(7), xmm0);    /* xmm0=h7h6h5h4h3h2h1h0 */ \
    movdqa_r2r(xmm6, xmm7);     /* make a copy */ \
402 403 404 405
    \
    punpcklwd_r2r(xmm0, xmm6);  /* xmm6=h3g3h3g2h1g1h0g0 */ \
    punpckhwd_r2r(xmm0, xmm7);  /* xmm7=h7g7h6g6h5g5h4g4 */ \
    \
406
    movdqa_r2r(xmm4, xmm3);     /* make a copy */ \
407 408 409
    punpckldq_r2r(xmm6, xmm4);  /* xmm4=h1g1f1e1h0g0f0e0 */ \
    \
    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 */ \
413 414 415
    punpckldq_r2r(xmm7, xmm5);  /* xmm5=h5g5f5e5h4g4f4e4 */ \
    \
    punpckhdq_r2r(xmm7, xmm6);  /* xmm6=h7g7f7e7h6g6f6e6 */ \
416
    movdqa_m2r(*I(0), xmm0);    /* xmm0=a7a6a5a4a3a2a1a0 */ \
417
    /* Free xmm7 */ \
418 419
    movdqa_m2r(*I(1), xmm1);    /* xmm1=b7b6b5b4b3b2b1b0 */ \
    movdqa_r2r(xmm0, xmm7);     /* make a copy */ \
420 421 422 423
    \
    punpcklwd_r2r(xmm1, xmm0);  /* xmm0=b3a3b2a2b1a1b0a0 */ \
    punpckhwd_r2r(xmm1, xmm7);  /* xmm7=b7a7b6a6b5a5b4a4 */ \
    /* Free xmm1 */ \
424 425
    movdqa_m2r(*I(2), xmm2);    /* xmm2=c7c6c5c4c3c2c1c0 */ \
    movdqa_m2r(*I(3), xmm3);    /* xmm3=d7d6d5d4d3d2d1d0 */ \
426
    \
427
    movdqa_r2r(xmm2, xmm1);     /* make a copy */ \
428 429 430
    punpcklwd_r2r(xmm3, xmm2);  /* xmm2=d3c3d2c2d1c1d0c0 */ \
    \
    punpckhwd_r2r(xmm3, xmm1);  /* xmm1=d7c7d6c6d5c5d4c4 */ \
431
    movdqa_r2r(xmm0, xmm3);     /* make a copy        */ \
432 433 434 435
    \
    punpckldq_r2r(xmm2, xmm0);  /* xmm0=d1c1b1a1d0c0b0a0 */ \
    punpckhdq_r2r(xmm2, xmm3);  /* xmm3=d3c3b3a3d2c2b2a2 */ \
    /* Free xmm2 */ \
436
    movdqa_r2r(xmm7, xmm2);     /* make a copy */ \
437 438 439
    punpckldq_r2r(xmm1, xmm2);  /* xmm2=d5c5b5a5d4c4b4a4 */ \
    \
    punpckhdq_r2r(xmm1, xmm7);  /* xmm7=d7c7b7a7d6c6b6a6 */ \
440
    movdqa_r2r(xmm0, xmm1);     /* make a copy */ \
441 442 443 444
    \
    punpcklqdq_r2r(xmm4, xmm0); /* xmm0=h0g0f0e0d0c0b0a0 */ \
    punpckhqdq_r2r(xmm4, xmm1); /* xmm1=h1g1g1e1d1c1b1a1 */ \
    \
445 446
    movdqa_r2m(xmm0, *I(0));    /* save I(0) */ \
    movdqa_r2m(xmm1, *I(1));    /* save I(1) */ \
447
    \
448 449
    movdqa_m2r(*I(6), xmm0);    /* load h3g3g3e3h2g2f2e2 */ \
    movdqa_r2r(xmm3, xmm1);     /* make a copy */ \
450 451 452 453
    \
    punpcklqdq_r2r(xmm0, xmm1); /* xmm1=h2g2f2e2d2c2b2a2 */ \
    punpckhqdq_r2r(xmm0, xmm3); /* xmm3=h3g3f3e3d3c3b3a3 */ \
    \
454
    movdqa_r2r(xmm2, xmm4);     /* make a copy */ \
455 456 457
    punpcklqdq_r2r(xmm5, xmm4); /* xmm4=h4g4f4e4d4c4b4a4 */ \
    \
    punpckhqdq_r2r(xmm5, xmm2); /* xmm2=h5g5f5e5d5c5b5a5 */ \
458
    movdqa_r2m(xmm1, *I(2));    /* save I(2) */ \
459
    \
460 461
    movdqa_r2m(xmm3, *I(3));    /* save I(3) */ \
    movdqa_r2m(xmm4, *I(4));    /* save I(4) */ \
462
    \
463 464
    movdqa_r2m(xmm2, *I(5));    /* save I(5) */ \
    movdqa_r2r(xmm7, xmm5);     /* make a copy */ \
465 466 467 468
    \
    punpcklqdq_r2r(xmm6, xmm5); /* xmm5=h6g6f6e6d6c6b6a6 */ \
    punpckhqdq_r2r(xmm6, xmm7); /* xmm7=h7g7f7e7d7c7b7a7 */ \
    \
469 470
    movdqa_r2m(xmm5, *I(6));    /* save I(6) */ \
    movdqa_r2m(xmm7, *I(7));    /* save I(7) */ \
471 472 473 474 475
    \
} /* End of Transpose Macro */


#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);     \
480 481 482 483
    \
    pmullw_m2r(*(ebx + 16), xmm1);     /* xmm1 = 17 16 15 14 13 12 11 10 */ \
    pshuflw_r2r(xmm0, xmm3, 0x078);    /* xmm3 = 07 06 05 04 01 03 02 00 */ \
    \
484 485
    movdqa_r2r(xmm1, xmm2);            /* xmm2 = 17 16 15 14 13 12 11 10 */ \
    movdqa_m2r(*(ecx), xmm7);          /* xmm7 = -- -- -- -- -- FF FF -- */ \
486
    \
487 488
    movdqa_m2r(*(eax + 32), xmm4);     \
    movdqa_m2r(*(eax + 64), xmm5);     \
489 490 491 492
    \
    pmullw_m2r(*(ebx + 32), xmm4);     /* xmm4 = 27 26 25 24 23 22 21 20 */ \
    pmullw_m2r(*(ebx + 64), xmm5);     /* xmm5 = 47 46 45 44 43 42 41 40 */ \
    \
493
    movdqa_m2r(*(ecx + 16), xmm6);     /* xmm6 = -- -- FF FF -- -- -- -- */ \
494 495 496 497 498 499 500 501 502 503 504
    pand_r2r(xmm2, xmm7);              /* xmm7 = -- -- -- -- -- 12 11 -- */ \
    \
    pand_r2r(xmm4, xmm6);              /* xmm6 = -- -- 25 24 -- -- -- -- */ \
    pxor_r2r(xmm7, xmm2);              /* xmm2 = 17 16 15 14 13 -- -- 10 */ \
    \
    pxor_r2r(xmm6, xmm4);              /* xmm4 = 27 26 -- -- 23 22 21 20 */ \
    pslldq_i2r(4, xmm7);               /* xmm7 = -- -- -- 12 11 -- -- -- */ \
    \
    pslldq_i2r(2, xmm6);               /* xmm6 = -- 25 24 -- -- -- -- -- */ \
    por_r2r(xmm6, xmm7);               /* xmm7 = -- 25 24 12 11 -- -- -- */ \
    \
505 506
    movdqa_m2r(*(ecx + 32), xmm0);     /* xmm0 = -- -- -- -- -- FF FF FF */ \
    movdqa_m2r(*(ecx + 48), xmm6);     /* xmm6 = -- -- -- -- FF -- -- -- */ \
507 508 509 510 511 512 513 514 515 516 517 518 519
    \
    pand_r2r(xmm3, xmm0);              /* xmm0 = -- -- -- -- -- 03 02 00 */ \
    pand_r2r(xmm5, xmm6);              /* xmm6 = -- -- -- -- 43 -- -- -- */ \
    \
    pxor_r2r(xmm0, xmm3);              /* xmm3 = 07 06 05 04 01 -- -- -- */ \
    pxor_r2r(xmm6, xmm5);              /* xmm5 = 47 46 45 44 -- 42 41 40 */ \
    \
    por_r2r(xmm7, xmm0);               /* xmm0 = -- 25 24 12 11 03 02 00 */ \
    pslldq_i2r(8, xmm6);               /* xmm6 = 43 -- -- -- -- -- -- -- */ \
    \
    por_r2r(xmm6, xmm0);               /* xmm0 = 43 25 24 12 11 03 02 00 */ \
    /* 02345 in use */ \
    \
520
    movdqa_m2r(*(ecx + 64 ), xmm1);    /* xmm1 = -- -- -- FF FF -- -- -- */ \
521 522
    pshuflw_r2r(xmm5, xmm5, 0x0B4);    /* xmm5 = 47 46 45 44 42 -- 41 40 */ \
    \
523 524
    movdqa_r2r(xmm1, xmm7);            /* xmm7 = -- -- -- FF FF -- -- -- */ \
    movdqa_r2r(xmm1, xmm6);            /* xmm6 = -- -- -- FF FF -- -- -- */ \
525
    \
526
    movdqa_r2m(xmm0, *(eax));          /* write  43 25 24 12 11 03 02 00 */ \
527 528 529 530 531 532 533 534 535
    pshufhw_r2r(xmm4, xmm4, 0x0C2);    /* xmm4 = 27 -- -- 26 23 22 21 20 */ \
    \
    pand_r2r(xmm4, xmm7);              /* xmm7 = -- -- -- 26 23 -- -- -- */ \
    pand_r2r(xmm5, xmm1);              /* xmm1 = -- -- -- 44 42 -- -- -- */ \
    \
    pxor_r2r(xmm7, xmm4);              /* xmm4 = 27 -- -- -- -- 22 21 20 */ \
    pxor_r2r(xmm1, xmm5);              /* xmm5 = 47 46 45 -- -- -- 41 40 */ \
    \
    pshuflw_r2r(xmm2, xmm2, 0x0C6);    /* xmm2 = 17 16 15 14 13 10 -- -- */ \
536
    movdqa_r2r(xmm6, xmm0);            /* xmm0 = -- -- -- FF FF -- -- -- */ \
537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556
    \
    pslldq_i2r(2, xmm7);               /* xmm7 = -- -- 26 23 -- -- -- -- */ \
    pslldq_i2r(6, xmm1);               /* xmm1 = 44 42 -- -- -- -- -- -- */ \
    \
    psrldq_i2r(2, xmm0);               /* xmm0 = -- -- -- -- FF FF -- -- */ \
    pand_r2r(xmm3, xmm6);              /* xmm6 = -- -- -- 04 01 -- -- -- */ \
    \
    pand_r2r(xmm2, xmm0);              /* xmm0 = -- -- -- -- 13 10 -- -- */ \
    pxor_r2r(xmm6, xmm3);              /* xmm3 = 07 06 05 -- -- -- -- -- */ \
    \
    pxor_r2r(xmm0, xmm2);              /* xmm2 = 17 16 15 14 -- -- -- -- */ \
    psrldq_i2r(6, xmm6);               /* xmm0 = -- -- -- -- -- -- 04 01 */ \
    \
    por_r2r(xmm7, xmm1);               /* xmm1 = 44 42 26 23 -- -- -- -- */ \
    por_r2r(xmm6, xmm0);               /* xmm1 = -- -- -- -- 13 10 04 01 */ \
    /* 12345 in use */    \
    por_r2r(xmm0, xmm1);               /* xmm1 = 44 42 26 23 13 10 04 01 */ \
    pshuflw_r2r(xmm4, xmm4, 0x093);    /* xmm4 = 27 -- -- -- 22 21 20 -- */ \
    \
    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 -- -- -- */ \
561 562 563 564 565 566
    \
    pand_r2r(xmm3, xmm0);              /* xmm0 = -- -- -- 06 -- -- -- -- */ \
    psrldq_i2r(12, xmm3);              /* xmm3 = -- -- -- -- -- -- 07 05 */ \
    \
    psrldq_i2r(8, xmm0);               /* xmm0 = -- -- -- -- -- -- -- 06 */ \
    \
567 568
    movdqa_m2r(*(ecx + 64), xmm6);     /* xmm6 = -- -- -- FF FF -- -- -- */ \
    movdqa_m2r(*(ecx + 96), xmm7);     /* xmm7 = -- -- -- -- FF FF -- -- */ \
569 570 571 572 573 574 575 576 577 578
    \
    pand_r2r(xmm4, xmm6);              /* xmm6 = -- -- -- 27 22 -- -- -- */ \
    pxor_r2r(xmm6, xmm4);              /* xmm4 = -- -- -- -- -- 21 20 -- */ \
    \
    por_r2r(xmm6, xmm3);               /* xmm3 = -- -- -- 27 22 -- 07 05 */ \
    pand_r2r(xmm4, xmm7);              /* xmm7 = -- -- -- -- -- 21 -- -- */ \
    \
    por_r2r(xmm7, xmm0);               /* xmm0 = -- -- -- -- -- 21 -- 06 */ \
    pxor_r2r(xmm7, xmm4);              /* xmm4 = -- -- -- -- -- -- 20 -- */ \
    \
579 580
    movdqa_m2r(*(ecx + 16 ), xmm6);    /* xmm6 = -- -- FF FF -- -- -- -- */ \
    movdqa_m2r(*(ecx + 64 ), xmm1);    /* xmm1 = -- -- -- FF FF -- -- -- */ \
581 582 583 584 585 586 587 588 589 590 591 592 593 594 595
    \
    pand_r2r(xmm2, xmm6);              /* xmm6 = -- -- 15 14 -- -- -- -- */ \
    pand_r2r(xmm6, xmm1);              /* xmm1 = -- -- -- 14 -- -- -- -- */ \
    \
    pxor_r2r(xmm6, xmm2);              /* xmm2 = 17 16 -- -- -- -- -- -- */ \
    pxor_r2r(xmm1, xmm6);              /* xmm6 = -- -- 15 -- -- -- -- -- */ \
    \
    psrldq_i2r(4, xmm1);               /* xmm1 = -- -- -- -- -- 14 -- -- */ \
    \
    psrldq_i2r(8, xmm6);               /* xmm6 = -- -- -- -- -- -- 15 -- */ \
    por_r2r(xmm1, xmm3);               /* xmm3 = -- -- -- 27 22 14 07 05 */ \
    \
    por_r2r(xmm6, xmm0);               /* xmm0 = -- -- -- -- -- 21 15 06 */ \
    pshufhw_r2r(xmm5, xmm5, 0x0E1);    /* xmm5 = 47 46 -- 45 -- -- 41 40 */ \
    \
596
    movdqa_m2r(*(ecx + 64), xmm1);     /* xmm1 = -- -- -- FF FF -- -- -- */ \
597 598
    pshuflw_r2r(xmm5, xmm5, 0x072);    /* xmm5 = 47 46 -- 45 41 -- 40 -- */ \
    \
599
    movdqa_r2r(xmm1, xmm6);            /* xmm6 = -- -- -- FF FF -- -- -- */ \
600 601 602 603 604 605 606 607
    pand_r2r(xmm5, xmm1);              /* xmm1 = -- -- -- 45 41 -- -- -- */ \
    \
    pxor_r2r(xmm1, xmm5);              /* xmm5 = 47 46 -- -- -- -- 40 -- */ \
    pslldq_i2r(4, xmm1);               /* xmm1 = -- 45 41 -- -- -- -- -- */ \
    \
    pshufd_r2r(xmm5, xmm5, 0x09C);     /* xmm5 = -- -- -- -- 47 46 40 -- */ \
    por_r2r(xmm1, xmm3);               /* xmm3 = -- 45 41 27 22 14 07 05 */ \
    \
608
    movdqa_m2r(*(eax + 96), xmm1);     /* xmm1 = 67 66 65 64 63 62 61 60 */ \
609 610
    pmullw_m2r(*(ebx + 96), xmm1);     \
    \
611
    movdqa_m2r(*(ecx), xmm7);          /* xmm7 = -- -- -- -- -- FF FF -- */ \
612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627
    \
    psrldq_i2r(8, xmm6);               /* xmm6 = -- -- -- -- -- -- -- FF */ \
    pand_r2r(xmm5, xmm7);              /* xmm7 = -- -- -- -- -- 46 40 -- */ \
    \
    pand_r2r(xmm1, xmm6);              /* xmm6 = -- -- -- -- -- -- -- 60 */ \
    pxor_r2r(xmm7, xmm5);              /* xmm5 = -- -- -- -- 47 -- -- -- */ \
    \
    pxor_r2r(xmm6, xmm1);              /* xmm1 = 67 66 65 64 63 62 61 -- */ \
    pslldq_i2r(2, xmm5);               /* xmm5 = -- -- -- 47 -- -- -- -- */ \
    \
    pslldq_i2r(14, xmm6);              /* xmm6 = 60 -- -- -- -- -- -- -- */ \
    por_r2r(xmm5, xmm4);               /* xmm4 = -- -- -- 47 -- -- 20 -- */ \
    \
    por_r2r(xmm6, xmm3);               /* xmm3 = 60 45 41 27 22 14 07 05 */ \
    pslldq_i2r(6, xmm7);               /* xmm7 = -- -- 46 40 -- -- -- -- */ \
    \
628
    movdqa_r2m(xmm3, *(eax+32));       /* write  60 45 41 27 22 14 07 05 */ \
629 630
    por_r2r(xmm7, xmm0);               /* xmm0 = -- -- 46 40 -- 21 15 06 */ \
    /* 0, 1, 2, 4 in use */    \
631 632
    movdqa_m2r(*(eax + 48), xmm3);     /* xmm3 = 37 36 35 34 33 32 31 30 */ \
    movdqa_m2r(*(eax + 80), xmm5);     /* xmm5 = 57 56 55 54 53 52 51 50 */ \
633 634 635 636
    \
    pmullw_m2r(*(ebx + 48), xmm3);     \
    pmullw_m2r(*(ebx + 80), xmm5);     \
    \
637 638
    movdqa_m2r(*(ecx + 64), xmm6);     /* xmm6 = -- -- -- FF FF -- -- -- */ \
    movdqa_m2r(*(ecx + 64), xmm7);     /* xmm7 = -- -- -- FF FF -- -- -- */ \
639 640 641 642 643 644 645 646 647 648 649 650 651 652
    \
    psrldq_i2r(8, xmm6);               /* xmm6 = -- -- -- -- -- -- -- FF */ \
    pslldq_i2r(8, xmm7);               /* xmm7 = FF -- -- -- -- -- -- -- */ \
    \
    pand_r2r(xmm3, xmm6);              /* xmm6 = -- -- -- -- -- -- -- 30 */ \
    pand_r2r(xmm5, xmm7);              /* xmm7 = 57 -- -- -- -- -- -- -- */ \
    \
    pxor_r2r(xmm6, xmm3);              /* xmm3 = 37 36 35 34 33 32 31 -- */ \
    pxor_r2r(xmm7, xmm5);              /* xmm5 = __ 56 55 54 53 52 51 50 */ \
    \
    pslldq_i2r(6, xmm6);               /* xmm6 = -- -- -- -- 30 -- -- -- */ \
    psrldq_i2r(2, xmm7);               /* xmm7 = -- 57 -- -- -- -- -- -- */ \
    \
    por_r2r(xmm7, xmm6);               /* xmm6 = -- 57 -- -- 30 -- -- -- */ \
653
    movdqa_m2r(*(ecx), xmm7);          /* xmm7 = -- -- -- -- -- FF FF -- */ \
654 655 656 657
    \
    por_r2r(xmm6, xmm0);               /* xmm0 = -- 57 46 40 30 21 15 06 */ \
    psrldq_i2r(2, xmm7);               /* xmm7 = -- -- -- -- -- -- FF FF */ \
    \
658
    movdqa_r2r(xmm2, xmm6);            /* xmm6 = 17 16 -- -- -- -- -- -- */ \
659 660 661 662 663 664 665 666 667 668 669 670
    pand_r2r(xmm1, xmm7);              /* xmm7 = -- -- -- -- -- -- 61 -- */ \
    \
    pslldq_i2r(2, xmm6);               /* xmm6 = 16 -- -- -- -- -- -- -- */ \
    psrldq_i2r(14, xmm2);              /* xmm2 = -- -- -- -- -- -- -- 17 */ \
    \
    pxor_r2r(xmm7, xmm1);              /* xmm1 = 67 66 65 64 63 62 -- -- */ \
    pslldq_i2r(12, xmm7);              /* xmm7 = 61 -- -- -- -- -- -- -- */ \
    \
    psrldq_i2r(14, xmm6);              /* xmm6 = -- -- -- -- -- -- -- 16 */ \
    por_r2r(xmm6, xmm4);               /* xmm4 = -- -- -- 47 -- -- 20 16 */ \
    \
    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 -- */ \
677 678
    pand_r2r(xmm3, xmm6);              /* xmm6 = -- -- -- -- -- -- 31 -- */ \
    \
679
    movdqa_r2r(xmm3, xmm7);            /* xmm7 = 37 36 35 34 33 32 31 -- */ \
680 681 682 683 684 685 686 687 688
    pxor_r2r(xmm6, xmm3);              /* xmm3 = 37 36 35 34 33 32 -- -- */ \
    \
    pslldq_i2r(2, xmm3);               /* xmm3 = 36 35 34 33 32 -- -- -- */ \
    pand_r2r(xmm1, xmm0);              /* xmm0 = -- -- -- -- -- 62 -- -- */ \
    \
    psrldq_i2r(14, xmm7);              /* xmm7 = -- -- -- -- -- -- -- 37 */ \
    pxor_r2r(xmm0, xmm1);              /* xmm1 = 67 66 65 64 63 -- -- -- */ \
    \
    por_r2r(xmm7, xmm6);               /* xmm6 = -- -- -- -- -- -- 31 37 */ \
689
    movdqa_m2r(*(ecx + 64), xmm7);     /* xmm7 = -- -- -- FF FF -- -- -- */ \
690 691 692 693 694 695 696 697 698 699 700 701 702 703 704
    \
    pshuflw_r2r(xmm6, xmm6, 0x01E);    /* xmm6 = -- -- -- -- 37 31 -- -- */ \
    pslldq_i2r(6, xmm7);               /* xmm7 = FF FF -- -- -- -- -- -- */ \
    \
    por_r2r(xmm6, xmm4);               /* xmm4 = -- -- -- 47 37 31 20 16 */ \
    pand_r2r(xmm5, xmm7);              /* xmm7 = -- 56 -- -- -- -- -- -- */ \
    \
    pslldq_i2r(8, xmm0);               /* xmm0 = -- 62 -- -- -- -- -- -- */ \
    pxor_r2r(xmm7, xmm5);              /* xmm5 = -- -- 55 54 53 52 51 50 */ \
    \
    psrldq_i2r(2, xmm7);               /* xmm7 = -- -- 56 -- -- -- -- -- */ \
    \
    pshufhw_r2r(xmm3, xmm3, 0x087);    /* xmm3 = 35 33 34 36 32 -- -- -- */ \
    por_r2r(xmm7, xmm0);               /* xmm0 = -- 62 56 -- -- -- -- -- */ \
    \
705
    movdqa_m2r(*(eax + 112), xmm7);    /* xmm7 = 77 76 75 74 73 72 71 70 */ \
706 707
    pmullw_m2r(*(ebx + 112), xmm7);     \
    \
708
    movdqa_m2r(*(ecx + 64), xmm6);     /* xmm6 = -- -- -- FF FF -- -- -- */ \
709 710 711 712 713
    por_r2r(xmm0, xmm4);               /* xmm4 = -- 62 56 47 37 31 20 16 */ \
    \
    pshuflw_r2r(xmm7, xmm7, 0x0E1);    /* xmm7 = 77 76 75 74 73 72 70 71 */ \
    psrldq_i2r(8, xmm6);               /* xmm6 = -- -- -- -- -- -- -- FF */ \
    \
714
    movdqa_m2r(*(ecx + 64), xmm0);     /* xmm0 = -- -- -- FF FF -- -- -- */ \
715 716 717 718 719 720 721 722 723 724 725 726
    pand_r2r(xmm7, xmm6);              /* xmm6 = -- -- -- -- -- -- -- 71 */ \
    \
    pand_r2r(xmm3, xmm0);              /* xmm0 = -- -- -- 36 32 -- -- -- */ \
    pxor_r2r(xmm6, xmm7);              /* xmm7 = 77 76 75 74 73 72 70 -- */ \
    \
    pxor_r2r(xmm0, xmm3);              /* xmm3 = 35 33 34 -- -- -- -- -- */ \
    pslldq_i2r(14, xmm6);              /* xmm6 = 71 -- -- -- -- -- -- -- */ \
    \
    psrldq_i2r(4, xmm0);               /* xmm0 = -- -- -- -- -- 36 32 -- */ \
    por_r2r(xmm6, xmm4);               /* xmm4 = 71 62 56 47 37 31 20 16 */ \
    \
    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 */ \
730 731
    pshufhw_r2r(xmm7, xmm7, 0x0D2);    /* xmm7 = 77 75 74 76 73 72 70 __ */ \
    \
732 733
    movdqa_m2r(*(ecx), xmm4);          /* xmm4 = -- -- -- -- -- FF FF -- */ \
    movdqa_m2r(*(ecx+48), xmm0);       /* xmm0 = -- -- -- -- FF -- -- -- */ \
734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755
    \
    pand_r2r(xmm5, xmm6);              /* xmm6 = -- -- 55 -- -- -- -- 50 */ \
    pand_r2r(xmm7, xmm4);              /* xmm4 = -- -- -- -- -- 72 70 -- */ \
    \
    pand_r2r(xmm1, xmm0);              /* xmm0 = -- -- -- -- 63 -- -- -- */ \
    pxor_r2r(xmm6, xmm5);              /* xmm5 = -- -- -- 54 53 52 51 -- */ \
    \
    pxor_r2r(xmm4, xmm7);              /* xmm7 = 77 75 74 76 73 -- -- -- */ \
    pxor_r2r(xmm0, xmm1);              /* xmm1 = 67 66 65 64 -- -- -- -- */ \
    \
    pshuflw_r2r(xmm6, xmm6, 0x02B);    /* xmm6 = -- -- 55 -- 50 -- -- -- */ \
    pslldq_i2r(10, xmm4);              /* xmm4 = 72 20 -- -- -- -- -- -- */ \
    \
    pshufhw_r2r(xmm6, xmm6, 0x0B1);    /* xmm6 = -- -- -- 55 50 -- -- -- */ \
    pslldq_i2r(4, xmm0);               /* xmm0 = -- -- 63 -- -- -- -- -- */ \
    \
    por_r2r(xmm4, xmm6);               /* xmm6 = 72 70 -- 55 50 -- -- -- */ \
    por_r2r(xmm0, xmm2);               /* xmm2 = -- -- 63 -- -- 36 32 17 */ \
    \
    por_r2r(xmm6, xmm2);               /* xmm2 = 72 70 64 55 50 36 32 17 */ \
    pshufhw_r2r(xmm1, xmm1, 0x0C9);    /* xmm1 = 67 64 66 65 -- -- -- -- */ \
    \
756 757
    movdqa_r2r(xmm3, xmm6);            /* xmm6 = 35 33 34 -- -- -- -- -- */ \
    movdqa_r2m(xmm2, *(eax+80));       /* write  72 70 64 55 50 36 32 17 */ \
758 759 760 761 762
    \
    psrldq_i2r(12, xmm6);              /* xmm6 = -- -- -- -- -- -- 35 33 */ \
    pslldq_i2r(4, xmm3);               /* xmm3 = 34 -- -- -- -- -- -- -- */ \
    \
    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 */ \
766 767 768 769 770
    psrldq_i2r(10, xmm7);              /* xmm7 = -- -- -- -- -- 77 75 74 */ \
    \
    pslldq_i2r(6, xmm4);               /* xmm4 = 76 73 -- -- -- -- -- -- */ \
    pslldq_i2r(12, xmm2);              /* xmm2 = 53 52 -- -- -- -- -- -- */ \
    \
771
    movdqa_r2r(xmm1, xmm0);            /* xmm0 = 67 64 66 65 -- -- -- -- */ \
772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794
    psrldq_i2r(12, xmm1);              /* xmm1 = -- -- -- -- -- -- 67 64 */ \
    \
    psrldq_i2r(6, xmm5);               /* xmm5 = -- -- -- -- -- -- 54 51 */ \
    psrldq_i2r(14, xmm3);              /* xmm3 = -- -- -- -- -- -- -- 34 */ \
    \
    pslldq_i2r(10, xmm7);              /* xmm7 = 77 75 74 -- -- -- -- -- */ \
    por_r2r(xmm6, xmm4);               /* xmm4 = 76 73 -- -- -- -- 35 33 */ \
    \
    psrldq_i2r(10, xmm2);              /* xmm2 = -- -- -- -- -- 53 52 -- */ \
    pslldq_i2r(4, xmm0);               /* xmm0 = 66 65 -- -- -- -- -- -- */ \
    \
    pslldq_i2r(8, xmm1);               /* xmm1 = -- -- 67 64 -- -- -- -- */ \
    por_r2r(xmm7, xmm3);               /* xmm3 = 77 75 74 -- -- -- -- 34 */ \
    \
    psrldq_i2r(6, xmm0);               /* xmm0 = -- -- -- 66 65 -- -- -- */ \
    pslldq_i2r(4, xmm5);               /* xmm5 = -- -- -- -- 54 51 -- -- */ \
    \
    por_r2r(xmm1, xmm4);               /* xmm4 = 76 73 67 64 -- -- 35 33 */ \
    por_r2r(xmm2, xmm3);               /* xmm3 = 77 75 74 -- -- 53 52 34 */ \
    \
    por_r2r(xmm5, xmm4);               /* xmm4 = 76 73 67 64 54 51 35 33 */ \
    por_r2r(xmm0, xmm3);               /* xmm3 = 77 75 74 66 65 53 52 34 */ \
    \
795 796
    movdqa_r2m(xmm4, *(eax+96));       /* write  76 73 67 64 54 51 35 33 */ \
    movdqa_r2m(xmm3, *(eax+112));      /* write  77 75 74 66 65 53 52 34 */ \
797 798 799 800
    \
} /* end of SSE2_Dequantize Macro */


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;
Reimar Döffinger's avatar
Reimar Döffinger committed
805 806
    const unsigned char *idct_data_bytes = (const unsigned char *)SSE2_idct_data;
    const unsigned char *Eight = (const unsigned char *)eight_data;
807 808

#define eax input_bytes
809
//#define ebx dequant_matrix_bytes
810 811 812 813 814 815 816
#define ecx dequant_const_bytes
#define edx idct_data_bytes

#define I(i) (eax + 16 * i)
#define O(i) (ebx + 16 * i)
#define C(i) (edx + 16 * (i-1))

817
 //   SSE2_Dequantize();
818 819 820 821 822 823 824

#undef ebx
#define ebx output_data_bytes

    SSE2_Row_IDCT();

    SSE2_Transpose();
825

826 827
    SSE2_Column_IDCT();
}
828 829 830 831 832 833 834 835 836 837 838 839

void ff_vp3_idct_put_sse2(uint8_t *dest, int line_size, DCTELEM *block)
{
    ff_vp3_idct_sse2(block);
    put_signed_pixels_clamped_mmx(block, dest, line_size);
}

void ff_vp3_idct_add_sse2(uint8_t *dest, int line_size, DCTELEM *block)
{
    ff_vp3_idct_sse2(block);
    add_pixels_clamped_mmx(block, dest, line_size);
}