2
* Copyright (C) 2004 the ffmpeg project
4
* This file is part of FFmpeg.
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.
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.
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
23
* SSE2-optimized functions cribbed from the original VP3 source code.
26
#include "../dsputil.h"
29
static DECLARE_ALIGNED_16(const unsigned short, SSE2_dequant_const[]) =
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
40
static DECLARE_ALIGNED_16(const unsigned int, eight_data[]) =
48
static DECLARE_ALIGNED_16(const unsigned short, SSE2_idct_data[7 * 8]) =
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
60
#define SSE2_Column_IDCT() { \
62
movdqu_m2r(*I(3), xmm2); /* xmm2 = i3 */ \
63
movdqu_m2r(*C(3), xmm6); /* xmm6 = c3 */ \
65
movdqu_r2r(xmm2, xmm4); /* xmm4 = i3 */ \
66
movdqu_m2r(*I(5), xmm7); /* xmm7 = i5 */ \
68
pmulhw_r2r(xmm6, xmm4); /* xmm4 = c3 * i3 - i3 */ \
69
movdqu_m2r(*C(5), xmm1); /* xmm1 = c5 */ \
71
pmulhw_r2r(xmm7, xmm6); /* xmm6 = c3 * i5 - i5 */ \
72
movdqu_r2r(xmm1, xmm5); /* xmm5 = c5 */ \
74
pmulhw_r2r(xmm2, xmm1); /* xmm1 = c5 * i3 - i3 */ \
75
movdqu_m2r(*I(1), xmm3); /* xmm3 = i1 */ \
77
pmulhw_r2r(xmm7, xmm5); /* xmm5 = c5 * i5 - i5 */ \
78
movdqu_m2r(*C(1), xmm0); /* xmm0 = c1 */ \
80
/* all registers are in use */ \
82
paddw_r2r(xmm2, xmm4); /* xmm4 = c3 * i3 */ \
83
paddw_r2r(xmm7, xmm6); /* xmm6 = c3 * i5 */ \
85
paddw_r2r(xmm1, xmm2); /* xmm2 = c5 * i3 */ \
86
movdqu_m2r(*I(7), xmm1); /* xmm1 = i7 */ \
88
paddw_r2r(xmm5, xmm7); /* xmm7 = c5 * i5 */ \
89
movdqu_r2r(xmm0, xmm5); /* xmm5 = c1 */ \
91
pmulhw_r2r(xmm3, xmm0); /* xmm0 = c1 * i1 - i1 */ \
92
paddsw_r2r(xmm7, xmm4); /* xmm4 = c3 * i3 + c5 * i5 = C */ \
94
pmulhw_r2r(xmm1, xmm5); /* xmm5 = c1 * i7 - i7 */ \
95
movdqu_m2r(*C(7), xmm7); /* xmm7 = c7 */ \
97
psubsw_r2r(xmm2, xmm6); /* xmm6 = c3 * i5 - c5 * i3 = D */ \
98
paddw_r2r(xmm3, xmm0); /* xmm0 = c1 * i1 */ \
100
pmulhw_r2r(xmm7, xmm3); /* xmm3 = c7 * i1 */ \
101
movdqu_m2r(*I(2), xmm2); /* xmm2 = i2 */ \
103
pmulhw_r2r(xmm1, xmm7); /* xmm7 = c7 * i7 */ \
104
paddw_r2r(xmm1, xmm5); /* xmm5 = c1 * i7 */ \
106
movdqu_r2r(xmm2, xmm1); /* xmm1 = i2 */ \
107
pmulhw_m2r(*C(2), xmm2); /* xmm2 = i2 * c2 -i2 */ \
109
psubsw_r2r(xmm5, xmm3); /* xmm3 = c7 * i1 - c1 * i7 = B */ \
110
movdqu_m2r(*I(6), xmm5); /* xmm5 = i6 */ \
112
paddsw_r2r(xmm7, xmm0); /* xmm0 = c1 * i1 + c7 * i7 = A */ \
113
movdqu_r2r(xmm5, xmm7); /* xmm7 = i6 */ \
115
psubsw_r2r(xmm4, xmm0); /* xmm0 = A - C */ \
116
pmulhw_m2r(*C(2), xmm5); /* xmm5 = c2 * i6 - i6 */ \
118
paddw_r2r(xmm1, xmm2); /* xmm2 = i2 * c2 */ \
119
pmulhw_m2r(*C(6), xmm1); /* xmm1 = c6 * i2 */ \
121
paddsw_r2r(xmm4, xmm4); /* xmm4 = C + C */ \
122
paddsw_r2r(xmm0, xmm4); /* xmm4 = A + C = C. */ \
124
psubsw_r2r(xmm6, xmm3); /* xmm3 = B - D */ \
125
paddw_r2r(xmm7, xmm5); /* xmm5 = c2 * i6 */ \
127
paddsw_r2r(xmm6, xmm6); /* xmm6 = D + D */ \
128
pmulhw_m2r(*C(6), xmm7); /* xmm7 = c6 * i6 */ \
130
paddsw_r2r(xmm3, xmm6); /* xmm6 = B + D = D. */ \
131
movdqu_r2m(xmm4, *I(1)); /* Save C. at I(1) */ \
133
psubsw_r2r(xmm5, xmm1); /* xmm1 = c6 * i2 - c2 * i6 = H */ \
134
movdqu_m2r(*C(4), xmm4); /* xmm4 = c4 */ \
136
movdqu_r2r(xmm3, xmm5); /* xmm5 = B - D */ \
137
pmulhw_r2r(xmm4, xmm3); /* xmm3 = ( c4 -1 ) * ( B - D ) */ \
139
paddsw_r2r(xmm2, xmm7); /* xmm7 = c2 * i2 + c6 * i6 = G */ \
140
movdqu_r2m(xmm6, *I(2)); /* Save D. at I(2) */ \
142
movdqu_r2r(xmm0, xmm2); /* xmm2 = A - C */ \
143
movdqu_m2r(*I(0), xmm6); /* xmm6 = i0 */ \
145
pmulhw_r2r(xmm4, xmm0); /* xmm0 = ( c4 - 1 ) * ( A - C ) = A. */ \
146
paddw_r2r(xmm3, xmm5); /* xmm5 = c4 * ( B - D ) = B. */ \
148
movdqu_m2r(*I(4), xmm3); /* xmm3 = i4 */ \
149
psubsw_r2r(xmm1, xmm5); /* xmm5 = B. - H = B.. */ \
151
paddw_r2r(xmm0, xmm2); /* xmm2 = c4 * ( A - C) = A. */ \
152
psubsw_r2r(xmm3, xmm6); /* xmm6 = i0 - i4 */ \
154
movdqu_r2r(xmm6, xmm0); /* xmm0 = i0 - i4 */ \
155
pmulhw_r2r(xmm4, xmm6); /* xmm6 = (c4 - 1) * (i0 - i4) = F */ \
157
paddsw_r2r(xmm3, xmm3); /* xmm3 = i4 + i4 */ \
158
paddsw_r2r(xmm1, xmm1); /* xmm1 = H + H */ \
160
paddsw_r2r(xmm0, xmm3); /* xmm3 = i0 + i4 */ \
161
paddsw_r2r(xmm5, xmm1); /* xmm1 = B. + H = H. */ \
163
pmulhw_r2r(xmm3, xmm4); /* xmm4 = ( c4 - 1 ) * ( i0 + i4 ) */ \
164
paddw_r2r(xmm0, xmm6); /* xmm6 = c4 * ( i0 - i4 ) */ \
166
psubsw_r2r(xmm2, xmm6); /* xmm6 = F - A. = F. */ \
167
paddsw_r2r(xmm2, xmm2); /* xmm2 = A. + A. */ \
169
movdqu_m2r(*I(1), xmm0); /* Load C. from I(1) */ \
170
paddsw_r2r(xmm6, xmm2); /* xmm2 = F + A. = A.. */ \
172
paddw_r2r(xmm3, xmm4); /* xmm4 = c4 * ( i0 + i4 ) = 3 */ \
173
psubsw_r2r(xmm1, xmm2); /* xmm2 = A.. - H. = R2 */ \
175
paddsw_m2r(*Eight, xmm2); /* Adjust R2 and R1 before shifting */ \
176
paddsw_r2r(xmm1, xmm1); /* xmm1 = H. + H. */ \
178
paddsw_r2r(xmm2, xmm1); /* xmm1 = A.. + H. = R1 */ \
179
psraw_i2r(4, xmm2); /* xmm2 = op2 */ \
181
psubsw_r2r(xmm7, xmm4); /* xmm4 = E - G = E. */ \
182
psraw_i2r(4, xmm1); /* xmm1 = op1 */ \
184
movdqu_m2r(*I(2), xmm3); /* Load D. from I(2) */ \
185
paddsw_r2r(xmm7, xmm7); /* xmm7 = G + G */ \
187
movdqu_r2m(xmm2, *O(2)); /* Write out op2 */ \
188
paddsw_r2r(xmm4, xmm7); /* xmm7 = E + G = G. */ \
190
movdqu_r2m(xmm1, *O(1)); /* Write out op1 */ \
191
psubsw_r2r(xmm3, xmm4); /* xmm4 = E. - D. = R4 */ \
193
paddsw_m2r(*Eight, xmm4); /* Adjust R4 and R3 before shifting */ \
194
paddsw_r2r(xmm3, xmm3); /* xmm3 = D. + D. */ \
196
paddsw_r2r(xmm4, xmm3); /* xmm3 = E. + D. = R3 */ \
197
psraw_i2r(4, xmm4); /* xmm4 = op4 */ \
199
psubsw_r2r(xmm5, xmm6); /* xmm6 = F. - B..= R6 */ \
200
psraw_i2r(4, xmm3); /* xmm3 = op3 */ \
202
paddsw_m2r(*Eight, xmm6); /* Adjust R6 and R5 before shifting */ \
203
paddsw_r2r(xmm5, xmm5); /* xmm5 = B.. + B.. */ \
205
paddsw_r2r(xmm6, xmm5); /* xmm5 = F. + B.. = R5 */ \
206
psraw_i2r(4, xmm6); /* xmm6 = op6 */ \
208
movdqu_r2m(xmm4, *O(4)); /* Write out op4 */ \
209
psraw_i2r(4, xmm5); /* xmm5 = op5 */ \
211
movdqu_r2m(xmm3, *O(3)); /* Write out op3 */ \
212
psubsw_r2r(xmm0, xmm7); /* xmm7 = G. - C. = R7 */ \
214
paddsw_m2r(*Eight, xmm7); /* Adjust R7 and R0 before shifting */ \
215
paddsw_r2r(xmm0, xmm0); /* xmm0 = C. + C. */ \
217
paddsw_r2r(xmm7, xmm0); /* xmm0 = G. + C. */ \
218
psraw_i2r(4, xmm7); /* xmm7 = op7 */ \
220
movdqu_r2m(xmm6, *O(6)); /* Write out op6 */ \
221
psraw_i2r(4, xmm0); /* xmm0 = op0 */ \
223
movdqu_r2m(xmm5, *O(5)); /* Write out op5 */ \
224
movdqu_r2m(xmm7, *O(7)); /* Write out op7 */ \
226
movdqu_r2m(xmm0, *O(0)); /* Write out op0 */ \
228
} /* End of SSE2_Column_IDCT macro */
231
#define SSE2_Row_IDCT() { \
233
movdqu_m2r(*I(3), xmm2); /* xmm2 = i3 */ \
234
movdqu_m2r(*C(3), xmm6); /* xmm6 = c3 */ \
236
movdqu_r2r(xmm2, xmm4); /* xmm4 = i3 */ \
237
movdqu_m2r(*I(5), xmm7); /* xmm7 = i5 */ \
239
pmulhw_r2r(xmm6, xmm4); /* xmm4 = c3 * i3 - i3 */ \
240
movdqu_m2r(*C(5), xmm1); /* xmm1 = c5 */ \
242
pmulhw_r2r(xmm7, xmm6); /* xmm6 = c3 * i5 - i5 */ \
243
movdqu_r2r(xmm1, xmm5); /* xmm5 = c5 */ \
245
pmulhw_r2r(xmm2, xmm1); /* xmm1 = c5 * i3 - i3 */ \
246
movdqu_m2r(*I(1), xmm3); /* xmm3 = i1 */ \
248
pmulhw_r2r(xmm7, xmm5); /* xmm5 = c5 * i5 - i5 */ \
249
movdqu_m2r(*C(1), xmm0); /* xmm0 = c1 */ \
251
/* all registers are in use */ \
253
paddw_r2r(xmm2, xmm4); /* xmm4 = c3 * i3 */ \
254
paddw_r2r(xmm7, xmm6); /* xmm6 = c3 * i5 */ \
256
paddw_r2r(xmm1, xmm2); /* xmm2 = c5 * i3 */ \
257
movdqu_m2r(*I(7), xmm1); /* xmm1 = i7 */ \
259
paddw_r2r(xmm5, xmm7); /* xmm7 = c5 * i5 */ \
260
movdqu_r2r(xmm0, xmm5); /* xmm5 = c1 */ \
262
pmulhw_r2r(xmm3, xmm0); /* xmm0 = c1 * i1 - i1 */ \
263
paddsw_r2r(xmm7, xmm4); /* xmm4 = c3 * i3 + c5 * i5 = C */ \
265
pmulhw_r2r(xmm1, xmm5); /* xmm5 = c1 * i7 - i7 */ \
266
movdqu_m2r(*C(7), xmm7); /* xmm7 = c7 */ \
268
psubsw_r2r(xmm2, xmm6); /* xmm6 = c3 * i5 - c5 * i3 = D */ \
269
paddw_r2r(xmm3, xmm0); /* xmm0 = c1 * i1 */ \
271
pmulhw_r2r(xmm7, xmm3); /* xmm3 = c7 * i1 */ \
272
movdqu_m2r(*I(2), xmm2); /* xmm2 = i2 */ \
274
pmulhw_r2r(xmm1, xmm7); /* xmm7 = c7 * i7 */ \
275
paddw_r2r(xmm1, xmm5); /* xmm5 = c1 * i7 */ \
277
movdqu_r2r(xmm2, xmm1); /* xmm1 = i2 */ \
278
pmulhw_m2r(*C(2), xmm2); /* xmm2 = i2 * c2 -i2 */ \
280
psubsw_r2r(xmm5, xmm3); /* xmm3 = c7 * i1 - c1 * i7 = B */ \
281
movdqu_m2r(*I(6), xmm5); /* xmm5 = i6 */ \
283
paddsw_r2r(xmm7, xmm0); /* xmm0 = c1 * i1 + c7 * i7 = A */ \
284
movdqu_r2r(xmm5, xmm7); /* xmm7 = i6 */ \
286
psubsw_r2r(xmm4, xmm0); /* xmm0 = A - C */ \
287
pmulhw_m2r(*C(2), xmm5); /* xmm5 = c2 * i6 - i6 */ \
289
paddw_r2r(xmm1, xmm2); /* xmm2 = i2 * c2 */ \
290
pmulhw_m2r(*C(6), xmm1); /* xmm1 = c6 * i2 */ \
292
paddsw_r2r(xmm4, xmm4); /* xmm4 = C + C */ \
293
paddsw_r2r(xmm0, xmm4); /* xmm4 = A + C = C. */ \
295
psubsw_r2r(xmm6, xmm3); /* xmm3 = B - D */ \
296
paddw_r2r(xmm7, xmm5); /* xmm5 = c2 * i6 */ \
298
paddsw_r2r(xmm6, xmm6); /* xmm6 = D + D */ \
299
pmulhw_m2r(*C(6), xmm7); /* xmm7 = c6 * i6 */ \
301
paddsw_r2r(xmm3, xmm6); /* xmm6 = B + D = D. */ \
302
movdqu_r2m(xmm4, *I(1)); /* Save C. at I(1) */ \
304
psubsw_r2r(xmm5, xmm1); /* xmm1 = c6 * i2 - c2 * i6 = H */ \
305
movdqu_m2r(*C(4), xmm4); /* xmm4 = c4 */ \
307
movdqu_r2r(xmm3, xmm5); /* xmm5 = B - D */ \
308
pmulhw_r2r(xmm4, xmm3); /* xmm3 = ( c4 -1 ) * ( B - D ) */ \
310
paddsw_r2r(xmm2, xmm7); /* xmm7 = c2 * i2 + c6 * i6 = G */ \
311
movdqu_r2m(xmm6, *I(2)); /* Save D. at I(2) */ \
313
movdqu_r2r(xmm0, xmm2); /* xmm2 = A - C */ \
314
movdqu_m2r(*I(0), xmm6); /* xmm6 = i0 */ \
316
pmulhw_r2r(xmm4, xmm0); /* xmm0 = ( c4 - 1 ) * ( A - C ) = A. */ \
317
paddw_r2r(xmm3, xmm5); /* xmm5 = c4 * ( B - D ) = B. */ \
319
movdqu_m2r(*I(4), xmm3); /* xmm3 = i4 */ \
320
psubsw_r2r(xmm1, xmm5); /* xmm5 = B. - H = B.. */ \
322
paddw_r2r(xmm0, xmm2); /* xmm2 = c4 * ( A - C) = A. */ \
323
psubsw_r2r(xmm3, xmm6); /* xmm6 = i0 - i4 */ \
325
movdqu_r2r(xmm6, xmm0); /* xmm0 = i0 - i4 */ \
326
pmulhw_r2r(xmm4, xmm6); /* xmm6 = ( c4 - 1 ) * ( i0 - i4 ) = F */ \
328
paddsw_r2r(xmm3, xmm3); /* xmm3 = i4 + i4 */ \
329
paddsw_r2r(xmm1, xmm1); /* xmm1 = H + H */ \
331
paddsw_r2r(xmm0, xmm3); /* xmm3 = i0 + i4 */ \
332
paddsw_r2r(xmm5, xmm1); /* xmm1 = B. + H = H. */ \
334
pmulhw_r2r(xmm3, xmm4); /* xmm4 = ( c4 - 1 ) * ( i0 + i4 ) */ \
335
paddw_r2r(xmm0, xmm6); /* xmm6 = c4 * ( i0 - i4 ) */ \
337
psubsw_r2r(xmm2, xmm6); /* xmm6 = F - A. = F. */ \
338
paddsw_r2r(xmm2, xmm2); /* xmm2 = A. + A. */ \
340
movdqu_m2r(*I(1), xmm0); /* Load C. from I(1) */ \
341
paddsw_r2r(xmm6, xmm2); /* xmm2 = F + A. = A.. */ \
343
paddw_r2r(xmm3, xmm4); /* xmm4 = c4 * ( i0 + i4 ) = 3 */ \
344
psubsw_r2r(xmm1, xmm2); /* xmm2 = A.. - H. = R2 */ \
346
paddsw_r2r(xmm1, xmm1); /* xmm1 = H. + H. */ \
347
paddsw_r2r(xmm2, xmm1); /* xmm1 = A.. + H. = R1 */ \
349
psubsw_r2r(xmm7, xmm4); /* xmm4 = E - G = E. */ \
351
movdqu_m2r(*I(2), xmm3); /* Load D. from I(2) */ \
352
paddsw_r2r(xmm7, xmm7); /* xmm7 = G + G */ \
354
movdqu_r2m(xmm2, *I(2)); /* Write out op2 */ \
355
paddsw_r2r(xmm4, xmm7); /* xmm7 = E + G = G. */ \
357
movdqu_r2m(xmm1, *I(1)); /* Write out op1 */ \
358
psubsw_r2r(xmm3, xmm4); /* xmm4 = E. - D. = R4 */ \
360
paddsw_r2r(xmm3, xmm3); /* xmm3 = D. + D. */ \
362
paddsw_r2r(xmm4, xmm3); /* xmm3 = E. + D. = R3 */ \
364
psubsw_r2r(xmm5, xmm6); /* xmm6 = F. - B..= R6 */ \
366
paddsw_r2r(xmm5, xmm5); /* xmm5 = B.. + B.. */ \
368
paddsw_r2r(xmm6, xmm5); /* xmm5 = F. + B.. = R5 */ \
370
movdqu_r2m(xmm4, *I(4)); /* Write out op4 */ \
372
movdqu_r2m(xmm3, *I(3)); /* Write out op3 */ \
373
psubsw_r2r(xmm0, xmm7); /* xmm7 = G. - C. = R7 */ \
375
paddsw_r2r(xmm0, xmm0); /* xmm0 = C. + C. */ \
377
paddsw_r2r(xmm7, xmm0); /* xmm0 = G. + C. */ \
379
movdqu_r2m(xmm6, *I(6)); /* Write out op6 */ \
381
movdqu_r2m(xmm5, *I(5)); /* Write out op5 */ \
382
movdqu_r2m(xmm7, *I(7)); /* Write out op7 */ \
384
movdqu_r2m(xmm0, *I(0)); /* Write out op0 */ \
386
} /* End of SSE2_Row_IDCT macro */
389
#define SSE2_Transpose() { \
391
movdqu_m2r(*I(4), xmm4); /* xmm4=e7e6e5e4e3e2e1e0 */ \
392
movdqu_m2r(*I(5), xmm0); /* xmm4=f7f6f5f4f3f2f1f0 */ \
394
movdqu_r2r(xmm4, xmm5); /* make a copy */ \
395
punpcklwd_r2r(xmm0, xmm4); /* xmm4=f3e3f2e2f1e1f0e0 */ \
397
punpckhwd_r2r(xmm0, xmm5); /* xmm5=f7e7f6e6f5e5f4e4 */ \
398
movdqu_m2r(*I(6), xmm6); /* xmm6=g7g6g5g4g3g2g1g0 */ \
400
movdqu_m2r(*I(7), xmm0); /* xmm0=h7h6h5h4h3h2h1h0 */ \
401
movdqu_r2r(xmm6, xmm7); /* make a copy */ \
403
punpcklwd_r2r(xmm0, xmm6); /* xmm6=h3g3h3g2h1g1h0g0 */ \
404
punpckhwd_r2r(xmm0, xmm7); /* xmm7=h7g7h6g6h5g5h4g4 */ \
406
movdqu_r2r(xmm4, xmm3); /* make a copy */ \
407
punpckldq_r2r(xmm6, xmm4); /* xmm4=h1g1f1e1h0g0f0e0 */ \
409
punpckhdq_r2r(xmm6, xmm3); /* xmm3=h3g3g3e3h2g2f2e2 */ \
410
movdqu_r2m(xmm3, *I(6)); /* save h3g3g3e3h2g2f2e2 */ \
412
movdqu_r2r(xmm5, xmm6); /* make a copy */ \
413
punpckldq_r2r(xmm7, xmm5); /* xmm5=h5g5f5e5h4g4f4e4 */ \
415
punpckhdq_r2r(xmm7, xmm6); /* xmm6=h7g7f7e7h6g6f6e6 */ \
416
movdqu_m2r(*I(0), xmm0); /* xmm0=a7a6a5a4a3a2a1a0 */ \
418
movdqu_m2r(*I(1), xmm1); /* xmm1=b7b6b5b4b3b2b1b0 */ \
419
movdqu_r2r(xmm0, xmm7); /* make a copy */ \
421
punpcklwd_r2r(xmm1, xmm0); /* xmm0=b3a3b2a2b1a1b0a0 */ \
422
punpckhwd_r2r(xmm1, xmm7); /* xmm7=b7a7b6a6b5a5b4a4 */ \
424
movdqu_m2r(*I(2), xmm2); /* xmm2=c7c6c5c4c3c2c1c0 */ \
425
movdqu_m2r(*I(3), xmm3); /* xmm3=d7d6d5d4d3d2d1d0 */ \
427
movdqu_r2r(xmm2, xmm1); /* make a copy */ \
428
punpcklwd_r2r(xmm3, xmm2); /* xmm2=d3c3d2c2d1c1d0c0 */ \
430
punpckhwd_r2r(xmm3, xmm1); /* xmm1=d7c7d6c6d5c5d4c4 */ \
431
movdqu_r2r(xmm0, xmm3); /* make a copy */ \
433
punpckldq_r2r(xmm2, xmm0); /* xmm0=d1c1b1a1d0c0b0a0 */ \
434
punpckhdq_r2r(xmm2, xmm3); /* xmm3=d3c3b3a3d2c2b2a2 */ \
436
movdqu_r2r(xmm7, xmm2); /* make a copy */ \
437
punpckldq_r2r(xmm1, xmm2); /* xmm2=d5c5b5a5d4c4b4a4 */ \
439
punpckhdq_r2r(xmm1, xmm7); /* xmm7=d7c7b7a7d6c6b6a6 */ \
440
movdqu_r2r(xmm0, xmm1); /* make a copy */ \
442
punpcklqdq_r2r(xmm4, xmm0); /* xmm0=h0g0f0e0d0c0b0a0 */ \
443
punpckhqdq_r2r(xmm4, xmm1); /* xmm1=h1g1g1e1d1c1b1a1 */ \
445
movdqu_r2m(xmm0, *I(0)); /* save I(0) */ \
446
movdqu_r2m(xmm1, *I(1)); /* save I(1) */ \
448
movdqu_m2r(*I(6), xmm0); /* load h3g3g3e3h2g2f2e2 */ \
449
movdqu_r2r(xmm3, xmm1); /* make a copy */ \
451
punpcklqdq_r2r(xmm0, xmm1); /* xmm1=h2g2f2e2d2c2b2a2 */ \
452
punpckhqdq_r2r(xmm0, xmm3); /* xmm3=h3g3f3e3d3c3b3a3 */ \
454
movdqu_r2r(xmm2, xmm4); /* make a copy */ \
455
punpcklqdq_r2r(xmm5, xmm4); /* xmm4=h4g4f4e4d4c4b4a4 */ \
457
punpckhqdq_r2r(xmm5, xmm2); /* xmm2=h5g5f5e5d5c5b5a5 */ \
458
movdqu_r2m(xmm1, *I(2)); /* save I(2) */ \
460
movdqu_r2m(xmm3, *I(3)); /* save I(3) */ \
461
movdqu_r2m(xmm4, *I(4)); /* save I(4) */ \
463
movdqu_r2m(xmm2, *I(5)); /* save I(5) */ \
464
movdqu_r2r(xmm7, xmm5); /* make a copy */ \
466
punpcklqdq_r2r(xmm6, xmm5); /* xmm5=h6g6f6e6d6c6b6a6 */ \
467
punpckhqdq_r2r(xmm6, xmm7); /* xmm7=h7g7f7e7d7c7b7a7 */ \
469
movdqu_r2m(xmm5, *I(6)); /* save I(6) */ \
470
movdqu_r2m(xmm7, *I(7)); /* save I(7) */ \
472
} /* End of Transpose Macro */
475
#define SSE2_Dequantize() { \
476
movdqu_m2r(*(eax), xmm0); \
478
pmullw_m2r(*(ebx), xmm0); /* xmm0 = 07 06 05 04 03 02 01 00 */ \
479
movdqu_m2r(*(eax + 16), xmm1); \
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 */ \
484
movdqu_r2r(xmm1, xmm2); /* xmm2 = 17 16 15 14 13 12 11 10 */ \
485
movdqu_m2r(*(ecx), xmm7); /* xmm7 = -- -- -- -- -- FF FF -- */ \
487
movdqu_m2r(*(eax + 32), xmm4); \
488
movdqu_m2r(*(eax + 64), xmm5); \
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 */ \
493
movdqu_m2r(*(ecx + 16), xmm6); /* xmm6 = -- -- FF FF -- -- -- -- */ \
494
pand_r2r(xmm2, xmm7); /* xmm7 = -- -- -- -- -- 12 11 -- */ \
496
pand_r2r(xmm4, xmm6); /* xmm6 = -- -- 25 24 -- -- -- -- */ \
497
pxor_r2r(xmm7, xmm2); /* xmm2 = 17 16 15 14 13 -- -- 10 */ \
499
pxor_r2r(xmm6, xmm4); /* xmm4 = 27 26 -- -- 23 22 21 20 */ \
500
pslldq_i2r(4, xmm7); /* xmm7 = -- -- -- 12 11 -- -- -- */ \
502
pslldq_i2r(2, xmm6); /* xmm6 = -- 25 24 -- -- -- -- -- */ \
503
por_r2r(xmm6, xmm7); /* xmm7 = -- 25 24 12 11 -- -- -- */ \
505
movdqu_m2r(*(ecx + 32), xmm0); /* xmm0 = -- -- -- -- -- FF FF FF */ \
506
movdqu_m2r(*(ecx + 48), xmm6); /* xmm6 = -- -- -- -- FF -- -- -- */ \
508
pand_r2r(xmm3, xmm0); /* xmm0 = -- -- -- -- -- 03 02 00 */ \
509
pand_r2r(xmm5, xmm6); /* xmm6 = -- -- -- -- 43 -- -- -- */ \
511
pxor_r2r(xmm0, xmm3); /* xmm3 = 07 06 05 04 01 -- -- -- */ \
512
pxor_r2r(xmm6, xmm5); /* xmm5 = 47 46 45 44 -- 42 41 40 */ \
514
por_r2r(xmm7, xmm0); /* xmm0 = -- 25 24 12 11 03 02 00 */ \
515
pslldq_i2r(8, xmm6); /* xmm6 = 43 -- -- -- -- -- -- -- */ \
517
por_r2r(xmm6, xmm0); /* xmm0 = 43 25 24 12 11 03 02 00 */ \
520
movdqu_m2r(*(ecx + 64 ), xmm1); /* xmm1 = -- -- -- FF FF -- -- -- */ \
521
pshuflw_r2r(xmm5, xmm5, 0x0B4); /* xmm5 = 47 46 45 44 42 -- 41 40 */ \
523
movdqu_r2r(xmm1, xmm7); /* xmm7 = -- -- -- FF FF -- -- -- */ \
524
movdqu_r2r(xmm1, xmm6); /* xmm6 = -- -- -- FF FF -- -- -- */ \
526
movdqu_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 */ \
529
pand_r2r(xmm4, xmm7); /* xmm7 = -- -- -- 26 23 -- -- -- */ \
530
pand_r2r(xmm5, xmm1); /* xmm1 = -- -- -- 44 42 -- -- -- */ \
532
pxor_r2r(xmm7, xmm4); /* xmm4 = 27 -- -- -- -- 22 21 20 */ \
533
pxor_r2r(xmm1, xmm5); /* xmm5 = 47 46 45 -- -- -- 41 40 */ \
535
pshuflw_r2r(xmm2, xmm2, 0x0C6); /* xmm2 = 17 16 15 14 13 10 -- -- */ \
536
movdqu_r2r(xmm6, xmm0); /* xmm0 = -- -- -- FF FF -- -- -- */ \
538
pslldq_i2r(2, xmm7); /* xmm7 = -- -- 26 23 -- -- -- -- */ \
539
pslldq_i2r(6, xmm1); /* xmm1 = 44 42 -- -- -- -- -- -- */ \
541
psrldq_i2r(2, xmm0); /* xmm0 = -- -- -- -- FF FF -- -- */ \
542
pand_r2r(xmm3, xmm6); /* xmm6 = -- -- -- 04 01 -- -- -- */ \
544
pand_r2r(xmm2, xmm0); /* xmm0 = -- -- -- -- 13 10 -- -- */ \
545
pxor_r2r(xmm6, xmm3); /* xmm3 = 07 06 05 -- -- -- -- -- */ \
547
pxor_r2r(xmm0, xmm2); /* xmm2 = 17 16 15 14 -- -- -- -- */ \
548
psrldq_i2r(6, xmm6); /* xmm0 = -- -- -- -- -- -- 04 01 */ \
550
por_r2r(xmm7, xmm1); /* xmm1 = 44 42 26 23 -- -- -- -- */ \
551
por_r2r(xmm6, xmm0); /* xmm1 = -- -- -- -- 13 10 04 01 */ \
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 -- */ \
556
pshufhw_r2r(xmm4, xmm4, 0x093); /* xmm4 = -- -- -- 27 22 21 20 -- */ \
557
movdqu_r2m(xmm1, *(eax + 16)); /* write 44 42 26 23 13 10 04 01 */ \
559
pshufhw_r2r(xmm3, xmm3, 0x0D2); /* xmm3 = 07 05 -- 06 -- -- -- -- */ \
560
movdqu_m2r(*(ecx + 64), xmm0); /* xmm0 = -- -- -- FF FF -- -- -- */ \
562
pand_r2r(xmm3, xmm0); /* xmm0 = -- -- -- 06 -- -- -- -- */ \
563
psrldq_i2r(12, xmm3); /* xmm3 = -- -- -- -- -- -- 07 05 */ \
565
psrldq_i2r(8, xmm0); /* xmm0 = -- -- -- -- -- -- -- 06 */ \
567
movdqu_m2r(*(ecx + 64), xmm6); /* xmm6 = -- -- -- FF FF -- -- -- */ \
568
movdqu_m2r(*(ecx + 96), xmm7); /* xmm7 = -- -- -- -- FF FF -- -- */ \
570
pand_r2r(xmm4, xmm6); /* xmm6 = -- -- -- 27 22 -- -- -- */ \
571
pxor_r2r(xmm6, xmm4); /* xmm4 = -- -- -- -- -- 21 20 -- */ \
573
por_r2r(xmm6, xmm3); /* xmm3 = -- -- -- 27 22 -- 07 05 */ \
574
pand_r2r(xmm4, xmm7); /* xmm7 = -- -- -- -- -- 21 -- -- */ \
576
por_r2r(xmm7, xmm0); /* xmm0 = -- -- -- -- -- 21 -- 06 */ \
577
pxor_r2r(xmm7, xmm4); /* xmm4 = -- -- -- -- -- -- 20 -- */ \
579
movdqu_m2r(*(ecx + 16 ), xmm6); /* xmm6 = -- -- FF FF -- -- -- -- */ \
580
movdqu_m2r(*(ecx + 64 ), xmm1); /* xmm1 = -- -- -- FF FF -- -- -- */ \
582
pand_r2r(xmm2, xmm6); /* xmm6 = -- -- 15 14 -- -- -- -- */ \
583
pand_r2r(xmm6, xmm1); /* xmm1 = -- -- -- 14 -- -- -- -- */ \
585
pxor_r2r(xmm6, xmm2); /* xmm2 = 17 16 -- -- -- -- -- -- */ \
586
pxor_r2r(xmm1, xmm6); /* xmm6 = -- -- 15 -- -- -- -- -- */ \
588
psrldq_i2r(4, xmm1); /* xmm1 = -- -- -- -- -- 14 -- -- */ \
590
psrldq_i2r(8, xmm6); /* xmm6 = -- -- -- -- -- -- 15 -- */ \
591
por_r2r(xmm1, xmm3); /* xmm3 = -- -- -- 27 22 14 07 05 */ \
593
por_r2r(xmm6, xmm0); /* xmm0 = -- -- -- -- -- 21 15 06 */ \
594
pshufhw_r2r(xmm5, xmm5, 0x0E1); /* xmm5 = 47 46 -- 45 -- -- 41 40 */ \
596
movdqu_m2r(*(ecx + 64), xmm1); /* xmm1 = -- -- -- FF FF -- -- -- */ \
597
pshuflw_r2r(xmm5, xmm5, 0x072); /* xmm5 = 47 46 -- 45 41 -- 40 -- */ \
599
movdqu_r2r(xmm1, xmm6); /* xmm6 = -- -- -- FF FF -- -- -- */ \
600
pand_r2r(xmm5, xmm1); /* xmm1 = -- -- -- 45 41 -- -- -- */ \
602
pxor_r2r(xmm1, xmm5); /* xmm5 = 47 46 -- -- -- -- 40 -- */ \
603
pslldq_i2r(4, xmm1); /* xmm1 = -- 45 41 -- -- -- -- -- */ \
605
pshufd_r2r(xmm5, xmm5, 0x09C); /* xmm5 = -- -- -- -- 47 46 40 -- */ \
606
por_r2r(xmm1, xmm3); /* xmm3 = -- 45 41 27 22 14 07 05 */ \
608
movdqu_m2r(*(eax + 96), xmm1); /* xmm1 = 67 66 65 64 63 62 61 60 */ \
609
pmullw_m2r(*(ebx + 96), xmm1); \
611
movdqu_m2r(*(ecx), xmm7); /* xmm7 = -- -- -- -- -- FF FF -- */ \
613
psrldq_i2r(8, xmm6); /* xmm6 = -- -- -- -- -- -- -- FF */ \
614
pand_r2r(xmm5, xmm7); /* xmm7 = -- -- -- -- -- 46 40 -- */ \
616
pand_r2r(xmm1, xmm6); /* xmm6 = -- -- -- -- -- -- -- 60 */ \
617
pxor_r2r(xmm7, xmm5); /* xmm5 = -- -- -- -- 47 -- -- -- */ \
619
pxor_r2r(xmm6, xmm1); /* xmm1 = 67 66 65 64 63 62 61 -- */ \
620
pslldq_i2r(2, xmm5); /* xmm5 = -- -- -- 47 -- -- -- -- */ \
622
pslldq_i2r(14, xmm6); /* xmm6 = 60 -- -- -- -- -- -- -- */ \
623
por_r2r(xmm5, xmm4); /* xmm4 = -- -- -- 47 -- -- 20 -- */ \
625
por_r2r(xmm6, xmm3); /* xmm3 = 60 45 41 27 22 14 07 05 */ \
626
pslldq_i2r(6, xmm7); /* xmm7 = -- -- 46 40 -- -- -- -- */ \
628
movdqu_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
movdqu_m2r(*(eax + 48), xmm3); /* xmm3 = 37 36 35 34 33 32 31 30 */ \
632
movdqu_m2r(*(eax + 80), xmm5); /* xmm5 = 57 56 55 54 53 52 51 50 */ \
634
pmullw_m2r(*(ebx + 48), xmm3); \
635
pmullw_m2r(*(ebx + 80), xmm5); \
637
movdqu_m2r(*(ecx + 64), xmm6); /* xmm6 = -- -- -- FF FF -- -- -- */ \
638
movdqu_m2r(*(ecx + 64), xmm7); /* xmm7 = -- -- -- FF FF -- -- -- */ \
640
psrldq_i2r(8, xmm6); /* xmm6 = -- -- -- -- -- -- -- FF */ \
641
pslldq_i2r(8, xmm7); /* xmm7 = FF -- -- -- -- -- -- -- */ \
643
pand_r2r(xmm3, xmm6); /* xmm6 = -- -- -- -- -- -- -- 30 */ \
644
pand_r2r(xmm5, xmm7); /* xmm7 = 57 -- -- -- -- -- -- -- */ \
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 */ \
649
pslldq_i2r(6, xmm6); /* xmm6 = -- -- -- -- 30 -- -- -- */ \
650
psrldq_i2r(2, xmm7); /* xmm7 = -- 57 -- -- -- -- -- -- */ \
652
por_r2r(xmm7, xmm6); /* xmm6 = -- 57 -- -- 30 -- -- -- */ \
653
movdqu_m2r(*(ecx), xmm7); /* xmm7 = -- -- -- -- -- FF FF -- */ \
655
por_r2r(xmm6, xmm0); /* xmm0 = -- 57 46 40 30 21 15 06 */ \
656
psrldq_i2r(2, xmm7); /* xmm7 = -- -- -- -- -- -- FF FF */ \
658
movdqu_r2r(xmm2, xmm6); /* xmm6 = 17 16 -- -- -- -- -- -- */ \
659
pand_r2r(xmm1, xmm7); /* xmm7 = -- -- -- -- -- -- 61 -- */ \
661
pslldq_i2r(2, xmm6); /* xmm6 = 16 -- -- -- -- -- -- -- */ \
662
psrldq_i2r(14, xmm2); /* xmm2 = -- -- -- -- -- -- -- 17 */ \
664
pxor_r2r(xmm7, xmm1); /* xmm1 = 67 66 65 64 63 62 -- -- */ \
665
pslldq_i2r(12, xmm7); /* xmm7 = 61 -- -- -- -- -- -- -- */ \
667
psrldq_i2r(14, xmm6); /* xmm6 = -- -- -- -- -- -- -- 16 */ \
668
por_r2r(xmm6, xmm4); /* xmm4 = -- -- -- 47 -- -- 20 16 */ \
670
por_r2r(xmm7, xmm0); /* xmm0 = 61 57 46 40 30 21 15 06 */ \
671
movdqu_m2r(*(ecx), xmm6); /* xmm6 = -- -- -- -- -- FF FF -- */ \
673
psrldq_i2r(2, xmm6); /* xmm6 = -- -- -- -- -- -- FF FF */ \
674
movdqu_r2m(xmm0, *(eax+48)); /* write 61 57 46 40 30 21 15 06 */ \
675
/* 1, 2, 3, 4, 5 in use */\
676
movdqu_m2r(*(ecx), xmm0); /* xmm0 = -- -- -- -- -- FF FF -- */ \
677
pand_r2r(xmm3, xmm6); /* xmm6 = -- -- -- -- -- -- 31 -- */ \
679
movdqu_r2r(xmm3, xmm7); /* xmm7 = 37 36 35 34 33 32 31 -- */ \
680
pxor_r2r(xmm6, xmm3); /* xmm3 = 37 36 35 34 33 32 -- -- */ \
682
pslldq_i2r(2, xmm3); /* xmm3 = 36 35 34 33 32 -- -- -- */ \
683
pand_r2r(xmm1, xmm0); /* xmm0 = -- -- -- -- -- 62 -- -- */ \
685
psrldq_i2r(14, xmm7); /* xmm7 = -- -- -- -- -- -- -- 37 */ \
686
pxor_r2r(xmm0, xmm1); /* xmm1 = 67 66 65 64 63 -- -- -- */ \
688
por_r2r(xmm7, xmm6); /* xmm6 = -- -- -- -- -- -- 31 37 */ \
689
movdqu_m2r(*(ecx + 64), xmm7); /* xmm7 = -- -- -- FF FF -- -- -- */ \
691
pshuflw_r2r(xmm6, xmm6, 0x01E); /* xmm6 = -- -- -- -- 37 31 -- -- */ \
692
pslldq_i2r(6, xmm7); /* xmm7 = FF FF -- -- -- -- -- -- */ \
694
por_r2r(xmm6, xmm4); /* xmm4 = -- -- -- 47 37 31 20 16 */ \
695
pand_r2r(xmm5, xmm7); /* xmm7 = -- 56 -- -- -- -- -- -- */ \
697
pslldq_i2r(8, xmm0); /* xmm0 = -- 62 -- -- -- -- -- -- */ \
698
pxor_r2r(xmm7, xmm5); /* xmm5 = -- -- 55 54 53 52 51 50 */ \
700
psrldq_i2r(2, xmm7); /* xmm7 = -- -- 56 -- -- -- -- -- */ \
702
pshufhw_r2r(xmm3, xmm3, 0x087); /* xmm3 = 35 33 34 36 32 -- -- -- */ \
703
por_r2r(xmm7, xmm0); /* xmm0 = -- 62 56 -- -- -- -- -- */ \
705
movdqu_m2r(*(eax + 112), xmm7); /* xmm7 = 77 76 75 74 73 72 71 70 */ \
706
pmullw_m2r(*(ebx + 112), xmm7); \
708
movdqu_m2r(*(ecx + 64), xmm6); /* xmm6 = -- -- -- FF FF -- -- -- */ \
709
por_r2r(xmm0, xmm4); /* xmm4 = -- 62 56 47 37 31 20 16 */ \
711
pshuflw_r2r(xmm7, xmm7, 0x0E1); /* xmm7 = 77 76 75 74 73 72 70 71 */ \
712
psrldq_i2r(8, xmm6); /* xmm6 = -- -- -- -- -- -- -- FF */ \
714
movdqu_m2r(*(ecx + 64), xmm0); /* xmm0 = -- -- -- FF FF -- -- -- */ \
715
pand_r2r(xmm7, xmm6); /* xmm6 = -- -- -- -- -- -- -- 71 */ \
717
pand_r2r(xmm3, xmm0); /* xmm0 = -- -- -- 36 32 -- -- -- */ \
718
pxor_r2r(xmm6, xmm7); /* xmm7 = 77 76 75 74 73 72 70 -- */ \
720
pxor_r2r(xmm0, xmm3); /* xmm3 = 35 33 34 -- -- -- -- -- */ \
721
pslldq_i2r(14, xmm6); /* xmm6 = 71 -- -- -- -- -- -- -- */ \
723
psrldq_i2r(4, xmm0); /* xmm0 = -- -- -- -- -- 36 32 -- */ \
724
por_r2r(xmm6, xmm4); /* xmm4 = 71 62 56 47 37 31 20 16 */ \
726
por_r2r(xmm0, xmm2); /* xmm2 = -- -- -- -- -- 36 32 17 */ \
727
movdqu_r2m(xmm4, *(eax + 64)); /* write 71 62 56 47 37 31 20 16 */ \
728
/* 1, 2, 3, 5, 7 in use */ \
729
movdqu_m2r(*(ecx + 80), xmm6); /* xmm6 = -- -- FF -- -- -- -- FF */ \
730
pshufhw_r2r(xmm7, xmm7, 0x0D2); /* xmm7 = 77 75 74 76 73 72 70 __ */ \
732
movdqu_m2r(*(ecx), xmm4); /* xmm4 = -- -- -- -- -- FF FF -- */ \
733
movdqu_m2r(*(ecx+48), xmm0); /* xmm0 = -- -- -- -- FF -- -- -- */ \
735
pand_r2r(xmm5, xmm6); /* xmm6 = -- -- 55 -- -- -- -- 50 */ \
736
pand_r2r(xmm7, xmm4); /* xmm4 = -- -- -- -- -- 72 70 -- */ \
738
pand_r2r(xmm1, xmm0); /* xmm0 = -- -- -- -- 63 -- -- -- */ \
739
pxor_r2r(xmm6, xmm5); /* xmm5 = -- -- -- 54 53 52 51 -- */ \
741
pxor_r2r(xmm4, xmm7); /* xmm7 = 77 75 74 76 73 -- -- -- */ \
742
pxor_r2r(xmm0, xmm1); /* xmm1 = 67 66 65 64 -- -- -- -- */ \
744
pshuflw_r2r(xmm6, xmm6, 0x02B); /* xmm6 = -- -- 55 -- 50 -- -- -- */ \
745
pslldq_i2r(10, xmm4); /* xmm4 = 72 20 -- -- -- -- -- -- */ \
747
pshufhw_r2r(xmm6, xmm6, 0x0B1); /* xmm6 = -- -- -- 55 50 -- -- -- */ \
748
pslldq_i2r(4, xmm0); /* xmm0 = -- -- 63 -- -- -- -- -- */ \
750
por_r2r(xmm4, xmm6); /* xmm6 = 72 70 -- 55 50 -- -- -- */ \
751
por_r2r(xmm0, xmm2); /* xmm2 = -- -- 63 -- -- 36 32 17 */ \
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 -- -- -- -- */ \
756
movdqu_r2r(xmm3, xmm6); /* xmm6 = 35 33 34 -- -- -- -- -- */ \
757
movdqu_r2m(xmm2, *(eax+80)); /* write 72 70 64 55 50 36 32 17 */ \
759
psrldq_i2r(12, xmm6); /* xmm6 = -- -- -- -- -- -- 35 33 */ \
760
pslldq_i2r(4, xmm3); /* xmm3 = 34 -- -- -- -- -- -- -- */ \
762
pshuflw_r2r(xmm5, xmm5, 0x04E); /* xmm5 = -- -- -- 54 51 -- 53 52 */ \
763
movdqu_r2r(xmm7, xmm4); /* xmm4 = 77 75 74 76 73 -- -- -- */ \
765
movdqu_r2r(xmm5, xmm2); /* xmm2 = -- -- -- 54 51 -- 53 52 */ \
766
psrldq_i2r(10, xmm7); /* xmm7 = -- -- -- -- -- 77 75 74 */ \
768
pslldq_i2r(6, xmm4); /* xmm4 = 76 73 -- -- -- -- -- -- */ \
769
pslldq_i2r(12, xmm2); /* xmm2 = 53 52 -- -- -- -- -- -- */ \
771
movdqu_r2r(xmm1, xmm0); /* xmm0 = 67 64 66 65 -- -- -- -- */ \
772
psrldq_i2r(12, xmm1); /* xmm1 = -- -- -- -- -- -- 67 64 */ \
774
psrldq_i2r(6, xmm5); /* xmm5 = -- -- -- -- -- -- 54 51 */ \
775
psrldq_i2r(14, xmm3); /* xmm3 = -- -- -- -- -- -- -- 34 */ \
777
pslldq_i2r(10, xmm7); /* xmm7 = 77 75 74 -- -- -- -- -- */ \
778
por_r2r(xmm6, xmm4); /* xmm4 = 76 73 -- -- -- -- 35 33 */ \
780
psrldq_i2r(10, xmm2); /* xmm2 = -- -- -- -- -- 53 52 -- */ \
781
pslldq_i2r(4, xmm0); /* xmm0 = 66 65 -- -- -- -- -- -- */ \
783
pslldq_i2r(8, xmm1); /* xmm1 = -- -- 67 64 -- -- -- -- */ \
784
por_r2r(xmm7, xmm3); /* xmm3 = 77 75 74 -- -- -- -- 34 */ \
786
psrldq_i2r(6, xmm0); /* xmm0 = -- -- -- 66 65 -- -- -- */ \
787
pslldq_i2r(4, xmm5); /* xmm5 = -- -- -- -- 54 51 -- -- */ \
789
por_r2r(xmm1, xmm4); /* xmm4 = 76 73 67 64 -- -- 35 33 */ \
790
por_r2r(xmm2, xmm3); /* xmm3 = 77 75 74 -- -- 53 52 34 */ \
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 */ \
795
movdqu_r2m(xmm4, *(eax+96)); /* write 76 73 67 64 54 51 35 33 */ \
796
movdqu_r2m(xmm3, *(eax+112)); /* write 77 75 74 66 65 53 52 34 */ \
798
} /* end of SSE2_Dequantize Macro */
801
void ff_vp3_idct_sse2(int16_t *input_data)
803
unsigned char *input_bytes = (unsigned char *)input_data;
804
unsigned char *output_data_bytes = (unsigned char *)input_data;
805
unsigned char *idct_data_bytes = (unsigned char *)SSE2_idct_data;
806
unsigned char *Eight = (unsigned char *)eight_data;
808
#define eax input_bytes
809
//#define ebx dequant_matrix_bytes
810
#define ecx dequant_const_bytes
811
#define edx idct_data_bytes
813
#define I(i) (eax + 16 * i)
814
#define O(i) (ebx + 16 * i)
815
#define C(i) (edx + 16 * (i-1))
817
// SSE2_Dequantize();
820
#define ebx output_data_bytes