2
* Copyright (C) 1999-2001 Aaron Holtzman <aholtzma@ess.engr.uvic.ca>
4
* This file is part of mpeg2dec, a free MPEG-2 video stream decoder.
5
* See http://libmpeg2.sourceforge.net/ for updates.
7
* mpeg2dec is free software; you can redistribute it and/or modify
8
* it under the terms of the GNU General Public License as published by
9
* the Free Software Foundation; either version 2 of the License, or
10
* (at your option) any later version.
12
* mpeg2dec is distributed in the hope that it will be useful,
13
* but WITHOUT ANY WARRANTY; without even the implied warranty of
14
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15
* GNU General Public License for more details.
17
* You should have received a copy of the GNU General Public License
18
* along with mpeg2dec; if not, write to the Free Software
19
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22
#include "libavutil/common.h"
23
#include "libavcodec/dsputil.h"
25
#include "libavutil/x86_cpu.h"
26
#include "dsputil_mmx.h"
31
#define round(bias) ((int)(((bias)+0.5) * (1<<ROW_SHIFT)))
32
#define rounder(bias) {round (bias), round (bias)}
36
/* C row IDCT - it is just here to document the MMXEXT and MMX versions */
37
static inline void idct_row (int16_t * row, int offset,
38
int16_t * table, int32_t * rounder)
40
int C1, C2, C3, C4, C5, C6, C7;
41
int a0, a1, a2, a3, b0, b1, b2, b3;
53
a0 = C4*row[0] + C2*row[2] + C4*row[4] + C6*row[6] + *rounder;
54
a1 = C4*row[0] + C6*row[2] - C4*row[4] - C2*row[6] + *rounder;
55
a2 = C4*row[0] - C6*row[2] - C4*row[4] + C2*row[6] + *rounder;
56
a3 = C4*row[0] - C2*row[2] + C4*row[4] - C6*row[6] + *rounder;
58
b0 = C1*row[1] + C3*row[3] + C5*row[5] + C7*row[7];
59
b1 = C3*row[1] - C7*row[3] - C1*row[5] - C5*row[7];
60
b2 = C5*row[1] - C1*row[3] + C7*row[5] + C3*row[7];
61
b3 = C7*row[1] - C5*row[3] + C3*row[5] - C1*row[7];
63
row[0] = (a0 + b0) >> ROW_SHIFT;
64
row[1] = (a1 + b1) >> ROW_SHIFT;
65
row[2] = (a2 + b2) >> ROW_SHIFT;
66
row[3] = (a3 + b3) >> ROW_SHIFT;
67
row[4] = (a3 - b3) >> ROW_SHIFT;
68
row[5] = (a2 - b2) >> ROW_SHIFT;
69
row[6] = (a1 - b1) >> ROW_SHIFT;
70
row[7] = (a0 - b0) >> ROW_SHIFT;
77
#define mmxext_table(c1,c2,c3,c4,c5,c6,c7) { c4, c2, -c4, -c2, \
86
static inline void mmxext_row_head (int16_t * const row, const int offset,
87
const int16_t * const table)
90
"movq (%0), %%mm2 \n\t" /* mm2 = x6 x4 x2 x0 */
92
"movq 8(%0), %%mm5 \n\t" /* mm5 = x7 x5 x3 x1 */
93
"movq %%mm2, %%mm0 \n\t" /* mm0 = x6 x4 x2 x0 */
95
"movq (%1), %%mm3 \n\t" /* mm3 = -C2 -C4 C2 C4 */
96
"movq %%mm5, %%mm6 \n\t" /* mm6 = x7 x5 x3 x1 */
98
"movq 8(%1), %%mm4 \n\t" /* mm4 = C6 C4 C6 C4 */
99
"pmaddwd %%mm0, %%mm3 \n\t" /* mm3 = -C4*x4-C2*x6 C4*x0+C2*x2 */
101
"pshufw $0x4e, %%mm2, %%mm2 \n\t" /* mm2 = x2 x0 x6 x4 */
102
:: "r" ((row+offset)), "r" (table)
106
static inline void mmxext_row (const int16_t * const table,
107
const int32_t * const rounder)
110
"movq 16(%0), %%mm1 \n\t" /* mm1 = -C5 -C1 C3 C1 */
111
"pmaddwd %%mm2, %%mm4 \n\t" /* mm4 = C4*x0+C6*x2 C4*x4+C6*x6 */
113
"pmaddwd 32(%0), %%mm0 \n\t" /* mm0 = C4*x4-C6*x6 C4*x0-C6*x2 */
114
"pshufw $0x4e, %%mm6, %%mm6 \n\t" /* mm6 = x3 x1 x7 x5 */
116
"movq 24(%0), %%mm7 \n\t" /* mm7 = -C7 C3 C7 C5 */
117
"pmaddwd %%mm5, %%mm1 \n\t" /* mm1= -C1*x5-C5*x7 C1*x1+C3*x3 */
119
"paddd (%1), %%mm3 \n\t" /* mm3 += rounder */
120
"pmaddwd %%mm6, %%mm7 \n\t" /* mm7 = C3*x1-C7*x3 C5*x5+C7*x7 */
122
"pmaddwd 40(%0), %%mm2 \n\t" /* mm2= C4*x0-C2*x2 -C4*x4+C2*x6 */
123
"paddd %%mm4, %%mm3 \n\t" /* mm3 = a1 a0 + rounder */
125
"pmaddwd 48(%0), %%mm5 \n\t" /* mm5 = C3*x5-C1*x7 C5*x1-C1*x3 */
126
"movq %%mm3, %%mm4 \n\t" /* mm4 = a1 a0 + rounder */
128
"pmaddwd 56(%0), %%mm6 \n\t" /* mm6 = C7*x1-C5*x3 C7*x5+C3*x7 */
129
"paddd %%mm7, %%mm1 \n\t" /* mm1 = b1 b0 */
131
"paddd (%1), %%mm0 \n\t" /* mm0 += rounder */
132
"psubd %%mm1, %%mm3 \n\t" /* mm3 = a1-b1 a0-b0 + rounder */
134
"psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm3 \n\t" /* mm3 = y6 y7 */
135
"paddd %%mm4, %%mm1 \n\t" /* mm1 = a1+b1 a0+b0 + rounder */
137
"paddd %%mm2, %%mm0 \n\t" /* mm0 = a3 a2 + rounder */
138
"psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm1 \n\t" /* mm1 = y1 y0 */
140
"paddd %%mm6, %%mm5 \n\t" /* mm5 = b3 b2 */
141
"movq %%mm0, %%mm4 \n\t" /* mm4 = a3 a2 + rounder */
143
"paddd %%mm5, %%mm0 \n\t" /* mm0 = a3+b3 a2+b2 + rounder */
144
"psubd %%mm5, %%mm4 \n\t" /* mm4 = a3-b3 a2-b2 + rounder */
145
: : "r" (table), "r" (rounder));
148
static inline void mmxext_row_tail (int16_t * const row, const int store)
151
"psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t" /* mm0 = y3 y2 */
153
"psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm4 \n\t" /* mm4 = y4 y5 */
155
"packssdw %%mm0, %%mm1 \n\t" /* mm1 = y3 y2 y1 y0 */
157
"packssdw %%mm3, %%mm4 \n\t" /* mm4 = y6 y7 y4 y5 */
159
"movq %%mm1, (%0) \n\t" /* save y3 y2 y1 y0 */
160
"pshufw $0xb1, %%mm4, %%mm4 \n\t" /* mm4 = y7 y6 y5 y4 */
164
"movq %%mm4, 8(%0) \n\t" /* save y7 y6 y5 y4 */
169
static inline void mmxext_row_mid (int16_t * const row, const int store,
171
const int16_t * const table)
174
"movq (%0,%1), %%mm2 \n\t" /* mm2 = x6 x4 x2 x0 */
175
"psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t" /* mm0 = y3 y2 */
177
"movq 8(%0,%1), %%mm5 \n\t" /* mm5 = x7 x5 x3 x1 */
178
"psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm4 \n\t" /* mm4 = y4 y5 */
180
"packssdw %%mm0, %%mm1 \n\t" /* mm1 = y3 y2 y1 y0 */
181
"movq %%mm5, %%mm6 \n\t" /* mm6 = x7 x5 x3 x1 */
183
"packssdw %%mm3, %%mm4 \n\t" /* mm4 = y6 y7 y4 y5 */
184
"movq %%mm2, %%mm0 \n\t" /* mm0 = x6 x4 x2 x0 */
186
"movq %%mm1, (%0,%2) \n\t" /* save y3 y2 y1 y0 */
187
"pshufw $0xb1, %%mm4, %%mm4\n\t" /* mm4 = y7 y6 y5 y4 */
189
"movq (%3), %%mm3 \n\t" /* mm3 = -C2 -C4 C2 C4 */
190
"movq %%mm4, 8(%0,%2) \n\t" /* save y7 y6 y5 y4 */
192
"pmaddwd %%mm0, %%mm3 \n\t" /* mm3= -C4*x4-C2*x6 C4*x0+C2*x2 */
194
"movq 8(%3), %%mm4 \n\t" /* mm4 = C6 C4 C6 C4 */
195
"pshufw $0x4e, %%mm2, %%mm2\n\t" /* mm2 = x2 x0 x6 x4 */
196
:: "r" (row), "r" ((x86_reg) (2*offset)), "r" ((x86_reg) (2*store)), "r" (table)
203
#define mmx_table(c1,c2,c3,c4,c5,c6,c7) { c4, c2, c4, c6, \
212
static inline void mmx_row_head (int16_t * const row, const int offset,
213
const int16_t * const table)
216
"movq (%0), %%mm2 \n\t" /* mm2 = x6 x4 x2 x0 */
218
"movq 8(%0), %%mm5 \n\t" /* mm5 = x7 x5 x3 x1 */
219
"movq %%mm2, %%mm0 \n\t" /* mm0 = x6 x4 x2 x0 */
221
"movq (%1), %%mm3 \n\t" /* mm3 = C6 C4 C2 C4 */
222
"movq %%mm5, %%mm6 \n\t" /* mm6 = x7 x5 x3 x1 */
224
"punpckldq %%mm0, %%mm0 \n\t" /* mm0 = x2 x0 x2 x0 */
226
"movq 8(%1), %%mm4 \n\t" /* mm4 = -C2 -C4 C6 C4 */
227
"pmaddwd %%mm0, %%mm3 \n\t" /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
229
"movq 16(%1), %%mm1 \n\t" /* mm1 = -C7 C3 C3 C1 */
230
"punpckhdq %%mm2, %%mm2 \n\t" /* mm2 = x6 x4 x6 x4 */
231
:: "r" ((row+offset)), "r" (table)
235
static inline void mmx_row (const int16_t * const table,
236
const int32_t * const rounder)
239
"pmaddwd %%mm2, %%mm4 \n\t" /* mm4 = -C4*x4-C2*x6 C4*x4+C6*x6 */
240
"punpckldq %%mm5, %%mm5 \n\t" /* mm5 = x3 x1 x3 x1 */
242
"pmaddwd 32(%0), %%mm0 \n\t" /* mm0 = C4*x0-C2*x2 C4*x0-C6*x2 */
243
"punpckhdq %%mm6, %%mm6 \n\t" /* mm6 = x7 x5 x7 x5 */
245
"movq 24(%0), %%mm7 \n\t" /* mm7 = -C5 -C1 C7 C5 */
246
"pmaddwd %%mm5, %%mm1 \n\t" /* mm1 = C3*x1-C7*x3 C1*x1+C3*x3 */
248
"paddd (%1), %%mm3 \n\t" /* mm3 += rounder */
249
"pmaddwd %%mm6, %%mm7 \n\t" /* mm7 = -C1*x5-C5*x7 C5*x5+C7*x7 */
251
"pmaddwd 40(%0), %%mm2 \n\t" /* mm2 = C4*x4-C6*x6 -C4*x4+C2*x6 */
252
"paddd %%mm4, %%mm3 \n\t" /* mm3 = a1 a0 + rounder */
254
"pmaddwd 48(%0), %%mm5 \n\t" /* mm5 = C7*x1-C5*x3 C5*x1-C1*x3 */
255
"movq %%mm3, %%mm4 \n\t" /* mm4 = a1 a0 + rounder */
257
"pmaddwd 56(%0), %%mm6 \n\t" /* mm6 = C3*x5-C1*x7 C7*x5+C3*x7 */
258
"paddd %%mm7, %%mm1 \n\t" /* mm1 = b1 b0 */
260
"paddd (%1), %%mm0 \n\t" /* mm0 += rounder */
261
"psubd %%mm1, %%mm3 \n\t" /* mm3 = a1-b1 a0-b0 + rounder */
263
"psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm3 \n\t" /* mm3 = y6 y7 */
264
"paddd %%mm4, %%mm1 \n\t" /* mm1 = a1+b1 a0+b0 + rounder */
266
"paddd %%mm2, %%mm0 \n\t" /* mm0 = a3 a2 + rounder */
267
"psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm1 \n\t" /* mm1 = y1 y0 */
269
"paddd %%mm6, %%mm5 \n\t" /* mm5 = b3 b2 */
270
"movq %%mm0, %%mm7 \n\t" /* mm7 = a3 a2 + rounder */
272
"paddd %%mm5, %%mm0 \n\t" /* mm0 = a3+b3 a2+b2 + rounder */
273
"psubd %%mm5, %%mm7 \n\t" /* mm7 = a3-b3 a2-b2 + rounder */
274
:: "r" (table), "r" (rounder)
278
static inline void mmx_row_tail (int16_t * const row, const int store)
281
"psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t" /* mm0 = y3 y2 */
283
"psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm7 \n\t" /* mm7 = y4 y5 */
285
"packssdw %%mm0, %%mm1 \n\t" /* mm1 = y3 y2 y1 y0 */
287
"packssdw %%mm3, %%mm7 \n\t" /* mm7 = y6 y7 y4 y5 */
289
"movq %%mm1, (%0) \n\t" /* save y3 y2 y1 y0 */
290
"movq %%mm7, %%mm4 \n\t" /* mm4 = y6 y7 y4 y5 */
292
"pslld $16, %%mm7 \n\t" /* mm7 = y7 0 y5 0 */
294
"psrld $16, %%mm4 \n\t" /* mm4 = 0 y6 0 y4 */
296
"por %%mm4, %%mm7 \n\t" /* mm7 = y7 y6 y5 y4 */
300
"movq %%mm7, 8(%0) \n\t" /* save y7 y6 y5 y4 */
305
static inline void mmx_row_mid (int16_t * const row, const int store,
306
const int offset, const int16_t * const table)
310
"movq (%0,%1), %%mm2 \n\t" /* mm2 = x6 x4 x2 x0 */
311
"psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t" /* mm0 = y3 y2 */
313
"movq 8(%0,%1), %%mm5 \n\t" /* mm5 = x7 x5 x3 x1 */
314
"psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm7 \n\t" /* mm7 = y4 y5 */
316
"packssdw %%mm0, %%mm1 \n\t" /* mm1 = y3 y2 y1 y0 */
317
"movq %%mm5, %%mm6 \n\t" /* mm6 = x7 x5 x3 x1 */
319
"packssdw %%mm3, %%mm7 \n\t" /* mm7 = y6 y7 y4 y5 */
320
"movq %%mm2, %%mm0 \n\t" /* mm0 = x6 x4 x2 x0 */
322
"movq %%mm1, (%0,%2) \n\t" /* save y3 y2 y1 y0 */
323
"movq %%mm7, %%mm1 \n\t" /* mm1 = y6 y7 y4 y5 */
325
"punpckldq %%mm0, %%mm0 \n\t" /* mm0 = x2 x0 x2 x0 */
326
"psrld $16, %%mm7 \n\t" /* mm7 = 0 y6 0 y4 */
328
"movq (%3), %%mm3 \n\t" /* mm3 = C6 C4 C2 C4 */
329
"pslld $16, %%mm1 \n\t" /* mm1 = y7 0 y5 0 */
331
"movq 8(%3), %%mm4 \n\t" /* mm4 = -C2 -C4 C6 C4 */
332
"por %%mm1, %%mm7 \n\t" /* mm7 = y7 y6 y5 y4 */
334
"movq 16(%3), %%mm1 \n\t" /* mm1 = -C7 C3 C3 C1 */
335
"punpckhdq %%mm2, %%mm2 \n\t" /* mm2 = x6 x4 x6 x4 */
337
"movq %%mm7, 8(%0,%2) \n\t" /* save y7 y6 y5 y4 */
338
"pmaddwd %%mm0, %%mm3 \n\t" /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
339
: : "r" (row), "r" ((x86_reg) (2*offset)), "r" ((x86_reg) (2*store)), "r" (table)
345
/* C column IDCT - it is just here to document the MMXEXT and MMX versions */
346
static inline void idct_col (int16_t * col, int offset)
348
/* multiplication - as implemented on mmx */
349
#define F(c,x) (((c) * (x)) >> 16)
351
/* saturation - it helps us handle torture test cases */
352
#define S(x) (((x)>32767) ? 32767 : ((x)<-32768) ? -32768 : (x))
354
int16_t x0, x1, x2, x3, x4, x5, x6, x7;
355
int16_t y0, y1, y2, y3, y4, y5, y6, y7;
356
int16_t a0, a1, a2, a3, b0, b1, b2, b3;
357
int16_t u04, v04, u26, v26, u17, v17, u35, v35, u12, v12;
372
u26 = S (F (T2, x6) + x2);
373
v26 = S (F (T2, x2) - x6);
380
u17 = S (F (T1, x7) + x1);
381
v17 = S (F (T1, x1) - x7);
382
u35 = S (F (T3, x5) + x3);
383
v35 = S (F (T3, x3) - x5);
389
u12 = S (2 * F (C4, u12));
390
v12 = S (2 * F (C4, v12));
394
y0 = S (a0 + b0) >> COL_SHIFT;
395
y1 = S (a1 + b1) >> COL_SHIFT;
396
y2 = S (a2 + b2) >> COL_SHIFT;
397
y3 = S (a3 + b3) >> COL_SHIFT;
399
y4 = S (a3 - b3) >> COL_SHIFT;
400
y5 = S (a2 - b2) >> COL_SHIFT;
401
y6 = S (a1 - b1) >> COL_SHIFT;
402
y7 = S (a0 - b0) >> COL_SHIFT;
416
/* MMX column IDCT */
417
static inline void idct_col (int16_t * const col, const int offset)
424
DECLARE_ALIGNED(8, static const short, t1_vector)[] = {
431
/* column code adapted from Peter Gubanov */
432
/* http://www.elecard.com/peter/idct.shtml */
435
"movq (%0), %%mm0 \n\t" /* mm0 = T1 */
437
"movq 2*8(%1), %%mm1 \n\t" /* mm1 = x1 */
438
"movq %%mm0, %%mm2 \n\t" /* mm2 = T1 */
440
"movq 7*2*8(%1), %%mm4 \n\t" /* mm4 = x7 */
441
"pmulhw %%mm1, %%mm0 \n\t" /* mm0 = T1*x1 */
443
"movq 16(%0), %%mm5 \n\t" /* mm5 = T3 */
444
"pmulhw %%mm4, %%mm2 \n\t" /* mm2 = T1*x7 */
446
"movq 2*5*8(%1), %%mm6 \n\t" /* mm6 = x5 */
447
"movq %%mm5, %%mm7 \n\t" /* mm7 = T3-1 */
449
"movq 3*8*2(%1), %%mm3 \n\t" /* mm3 = x3 */
450
"psubsw %%mm4, %%mm0 \n\t" /* mm0 = v17 */
452
"movq 8(%0), %%mm4 \n\t" /* mm4 = T2 */
453
"pmulhw %%mm3, %%mm5 \n\t" /* mm5 = (T3-1)*x3 */
455
"paddsw %%mm2, %%mm1 \n\t" /* mm1 = u17 */
456
"pmulhw %%mm6, %%mm7 \n\t" /* mm7 = (T3-1)*x5 */
460
"movq %%mm4, %%mm2 \n\t" /* mm2 = T2 */
461
"paddsw %%mm3, %%mm5 \n\t" /* mm5 = T3*x3 */
463
"pmulhw 2*8*2(%1), %%mm4 \n\t" /* mm4 = T2*x2 */
464
"paddsw %%mm6, %%mm7 \n\t" /* mm7 = T3*x5 */
466
"psubsw %%mm6, %%mm5 \n\t" /* mm5 = v35 */
467
"paddsw %%mm3, %%mm7 \n\t" /* mm7 = u35 */
469
"movq 6*8*2(%1), %%mm3 \n\t" /* mm3 = x6 */
470
"movq %%mm0, %%mm6 \n\t" /* mm6 = v17 */
472
"pmulhw %%mm3, %%mm2 \n\t" /* mm2 = T2*x6 */
473
"psubsw %%mm5, %%mm0 \n\t" /* mm0 = b3 */
475
"psubsw %%mm3, %%mm4 \n\t" /* mm4 = v26 */
476
"paddsw %%mm6, %%mm5 \n\t" /* mm5 = v12 */
478
"movq %%mm0, 3*8*2(%1)\n\t" /* save b3 in scratch0 */
479
"movq %%mm1, %%mm6 \n\t" /* mm6 = u17 */
481
"paddsw 2*8*2(%1), %%mm2 \n\t" /* mm2 = u26 */
482
"paddsw %%mm7, %%mm6 \n\t" /* mm6 = b0 */
484
"psubsw %%mm7, %%mm1 \n\t" /* mm1 = u12 */
485
"movq %%mm1, %%mm7 \n\t" /* mm7 = u12 */
487
"movq 0*8(%1), %%mm3 \n\t" /* mm3 = x0 */
488
"paddsw %%mm5, %%mm1 \n\t" /* mm1 = u12+v12 */
490
"movq 24(%0), %%mm0 \n\t" /* mm0 = C4/2 */
491
"psubsw %%mm5, %%mm7 \n\t" /* mm7 = u12-v12 */
493
"movq %%mm6, 5*8*2(%1)\n\t" /* save b0 in scratch1 */
494
"pmulhw %%mm0, %%mm1 \n\t" /* mm1 = b1/2 */
496
"movq %%mm4, %%mm6 \n\t" /* mm6 = v26 */
497
"pmulhw %%mm0, %%mm7 \n\t" /* mm7 = b2/2 */
499
"movq 4*8*2(%1), %%mm5 \n\t" /* mm5 = x4 */
500
"movq %%mm3, %%mm0 \n\t" /* mm0 = x0 */
502
"psubsw %%mm5, %%mm3 \n\t" /* mm3 = v04 */
503
"paddsw %%mm5, %%mm0 \n\t" /* mm0 = u04 */
505
"paddsw %%mm3, %%mm4 \n\t" /* mm4 = a1 */
506
"movq %%mm0, %%mm5 \n\t" /* mm5 = u04 */
508
"psubsw %%mm6, %%mm3 \n\t" /* mm3 = a2 */
509
"paddsw %%mm2, %%mm5 \n\t" /* mm5 = a0 */
511
"paddsw %%mm1, %%mm1 \n\t" /* mm1 = b1 */
512
"psubsw %%mm2, %%mm0 \n\t" /* mm0 = a3 */
514
"paddsw %%mm7, %%mm7 \n\t" /* mm7 = b2 */
515
"movq %%mm3, %%mm2 \n\t" /* mm2 = a2 */
517
"movq %%mm4, %%mm6 \n\t" /* mm6 = a1 */
518
"paddsw %%mm7, %%mm3 \n\t" /* mm3 = a2+b2 */
520
"psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm3\n\t" /* mm3 = y2 */
521
"paddsw %%mm1, %%mm4\n\t" /* mm4 = a1+b1 */
523
"psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm4\n\t" /* mm4 = y1 */
524
"psubsw %%mm1, %%mm6 \n\t" /* mm6 = a1-b1 */
526
"movq 5*8*2(%1), %%mm1 \n\t" /* mm1 = b0 */
527
"psubsw %%mm7, %%mm2 \n\t" /* mm2 = a2-b2 */
529
"psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm6\n\t" /* mm6 = y6 */
530
"movq %%mm5, %%mm7 \n\t" /* mm7 = a0 */
532
"movq %%mm4, 1*8*2(%1)\n\t" /* save y1 */
533
"psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm2\n\t" /* mm2 = y5 */
535
"movq %%mm3, 2*8*2(%1)\n\t" /* save y2 */
536
"paddsw %%mm1, %%mm5 \n\t" /* mm5 = a0+b0 */
538
"movq 3*8*2(%1), %%mm4 \n\t" /* mm4 = b3 */
539
"psubsw %%mm1, %%mm7 \n\t" /* mm7 = a0-b0 */
541
"psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm5\n\t" /* mm5 = y0 */
542
"movq %%mm0, %%mm3 \n\t" /* mm3 = a3 */
544
"movq %%mm2, 5*8*2(%1)\n\t" /* save y5 */
545
"psubsw %%mm4, %%mm3 \n\t" /* mm3 = a3-b3 */
547
"psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm7\n\t" /* mm7 = y7 */
548
"paddsw %%mm0, %%mm4 \n\t" /* mm4 = a3+b3 */
550
"movq %%mm5, 0*8*2(%1)\n\t" /* save y0 */
551
"psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm3\n\t" /* mm3 = y4 */
553
"movq %%mm6, 6*8*2(%1)\n\t" /* save y6 */
554
"psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm4\n\t" /* mm4 = y3 */
556
"movq %%mm7, 7*8*2(%1)\n\t" /* save y7 */
558
"movq %%mm3, 4*8*2(%1)\n\t" /* save y4 */
560
"movq %%mm4, 3*8*2(%1)\n\t" /* save y3 */
561
:: "r" (t1_vector), "r" (col+offset)
571
DECLARE_ALIGNED(8, static const int32_t, rounder0)[] =
572
rounder ((1 << (COL_SHIFT - 1)) - 0.5);
573
DECLARE_ALIGNED(8, static const int32_t, rounder4)[] = rounder (0);
574
DECLARE_ALIGNED(8, static const int32_t, rounder1)[] =
575
rounder (1.25683487303); /* C1*(C1/C4+C1+C7)/2 */
576
DECLARE_ALIGNED(8, static const int32_t, rounder7)[] =
577
rounder (-0.25); /* C1*(C7/C4+C7-C1)/2 */
578
DECLARE_ALIGNED(8, static const int32_t, rounder2)[] =
579
rounder (0.60355339059); /* C2 * (C6+C2)/2 */
580
DECLARE_ALIGNED(8, static const int32_t, rounder6)[] =
581
rounder (-0.25); /* C2 * (C6-C2)/2 */
582
DECLARE_ALIGNED(8, static const int32_t, rounder3)[] =
583
rounder (0.087788325588); /* C3*(-C3/C4+C3+C5)/2 */
584
DECLARE_ALIGNED(8, static const int32_t, rounder5)[] =
585
rounder (-0.441341716183); /* C3*(-C5/C4+C5-C3)/2 */
590
#define declare_idct(idct,table,idct_row_head,idct_row,idct_row_tail,idct_row_mid) \
591
void idct (int16_t * const block) \
593
DECLARE_ALIGNED(16, static const int16_t, table04)[] = \
594
table (22725, 21407, 19266, 16384, 12873, 8867, 4520); \
595
DECLARE_ALIGNED(16, static const int16_t, table17)[] = \
596
table (31521, 29692, 26722, 22725, 17855, 12299, 6270); \
597
DECLARE_ALIGNED(16, static const int16_t, table26)[] = \
598
table (29692, 27969, 25172, 21407, 16819, 11585, 5906); \
599
DECLARE_ALIGNED(16, static const int16_t, table35)[] = \
600
table (26722, 25172, 22654, 19266, 15137, 10426, 5315); \
602
idct_row_head (block, 0*8, table04); \
603
idct_row (table04, rounder0); \
604
idct_row_mid (block, 0*8, 4*8, table04); \
605
idct_row (table04, rounder4); \
606
idct_row_mid (block, 4*8, 1*8, table17); \
607
idct_row (table17, rounder1); \
608
idct_row_mid (block, 1*8, 7*8, table17); \
609
idct_row (table17, rounder7); \
610
idct_row_mid (block, 7*8, 2*8, table26); \
611
idct_row (table26, rounder2); \
612
idct_row_mid (block, 2*8, 6*8, table26); \
613
idct_row (table26, rounder6); \
614
idct_row_mid (block, 6*8, 3*8, table35); \
615
idct_row (table35, rounder3); \
616
idct_row_mid (block, 3*8, 5*8, table35); \
617
idct_row (table35, rounder5); \
618
idct_row_tail (block, 5*8); \
620
idct_col (block, 0); \
621
idct_col (block, 4); \
624
declare_idct (ff_mmxext_idct, mmxext_table,
625
mmxext_row_head, mmxext_row, mmxext_row_tail, mmxext_row_mid)
627
declare_idct (ff_mmx_idct, mmx_table,
628
mmx_row_head, mmx_row, mmx_row_tail, mmx_row_mid)