Revision f33802a0

View differences:

libavcodec/i386/idct_mmx.c
1
/*
2
 * Note: For libavcodec, this code can also be used under the LGPL license
3
 */
4
/*
5
 * idct_mmx.c
6
 * Copyright (C) 1999-2001 Aaron Holtzman <aholtzma@ess.engr.uvic.ca>
7
 *
8
 * This file is part of mpeg2dec, a free MPEG-2 video stream decoder.
9
 *
10
 * mpeg2dec is free software; you can redistribute it and/or modify
11
 * it under the terms of the GNU General Public License as published by
12
 * the Free Software Foundation; either version 2 of the License, or
13
 * (at your option) any later version.
14
 *
15
 * mpeg2dec is distributed in the hope that it will be useful,
16
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
18
 * GNU General Public License for more details.
19
 *
20
 * You should have received a copy of the GNU General Public License
21
 * along with this program; if not, write to the Free Software
22
 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
23
 */
24

  
25
#include <inttypes.h>
26

  
27
#include "mmx.h"
28

  
29
#define ATTR_ALIGN(align) __attribute__ ((__aligned__ (align)))
30

  
31
#define ROW_SHIFT 11
32
#define COL_SHIFT 6
33

  
34
#define round(bias) ((int)(((bias)+0.5) * (1<<ROW_SHIFT)))
35
#define rounder(bias) {round (bias), round (bias)}
36

  
37
#if 0
38
/* C row IDCT - its just here to document the MMXEXT and MMX versions */
39
static inline void idct_row (int16_t * row, int offset,
40
			     int16_t * table, int32_t * rounder)
41
{
42
    int C1, C2, C3, C4, C5, C6, C7;
43
    int a0, a1, a2, a3, b0, b1, b2, b3;
44

  
45
    row += offset;
46

  
47
    C1 = table[1];
48
    C2 = table[2];
49
    C3 = table[3];
50
    C4 = table[4];
51
    C5 = table[5];
52
    C6 = table[6];
53
    C7 = table[7];
54

  
55
    a0 = C4*row[0] + C2*row[2] + C4*row[4] + C6*row[6] + *rounder;
56
    a1 = C4*row[0] + C6*row[2] - C4*row[4] - C2*row[6] + *rounder;
57
    a2 = C4*row[0] - C6*row[2] - C4*row[4] + C2*row[6] + *rounder;
58
    a3 = C4*row[0] - C2*row[2] + C4*row[4] - C6*row[6] + *rounder;
59

  
60
    b0 = C1*row[1] + C3*row[3] + C5*row[5] + C7*row[7];
61
    b1 = C3*row[1] - C7*row[3] - C1*row[5] - C5*row[7];
62
    b2 = C5*row[1] - C1*row[3] + C7*row[5] + C3*row[7];
63
    b3 = C7*row[1] - C5*row[3] + C3*row[5] - C1*row[7];
64

  
65
    row[0] = (a0 + b0) >> ROW_SHIFT;
66
    row[1] = (a1 + b1) >> ROW_SHIFT;
67
    row[2] = (a2 + b2) >> ROW_SHIFT;
68
    row[3] = (a3 + b3) >> ROW_SHIFT;
69
    row[4] = (a3 - b3) >> ROW_SHIFT;
70
    row[5] = (a2 - b2) >> ROW_SHIFT;
71
    row[6] = (a1 - b1) >> ROW_SHIFT;
72
    row[7] = (a0 - b0) >> ROW_SHIFT;
73
}
74
#endif
75

  
76

  
77
/* MMXEXT row IDCT */
78

  
79
#define mmxext_table(c1,c2,c3,c4,c5,c6,c7)	{  c4,  c2, -c4, -c2,	\
80
						   c4,  c6,  c4,  c6,	\
81
						   c1,  c3, -c1, -c5,	\
82
						   c5,  c7,  c3, -c7,	\
83
						   c4, -c6,  c4, -c6,	\
84
						  -c4,  c2,  c4, -c2,	\
85
						   c5, -c1,  c3, -c1,	\
86
						   c7,  c3,  c7, -c5 }
87

  
88
static inline void mmxext_row_head (int16_t * row, int offset, int16_t * table)
89
{
90
    movq_m2r (*(row+offset), mm2);	// mm2 = x6 x4 x2 x0
91

  
92
    movq_m2r (*(row+offset+4), mm5);	// mm5 = x7 x5 x3 x1
93
    movq_r2r (mm2, mm0);		// mm0 = x6 x4 x2 x0
94

  
95
    movq_m2r (*table, mm3);		// mm3 = -C2 -C4 C2 C4
96
    movq_r2r (mm5, mm6);		// mm6 = x7 x5 x3 x1
97

  
98
    movq_m2r (*(table+4), mm4);		// mm4 = C6 C4 C6 C4
99
    pmaddwd_r2r (mm0, mm3);		// mm3 = -C4*x4-C2*x6 C4*x0+C2*x2
100

  
101
    pshufw_r2r (mm2, mm2, 0x4e);	// mm2 = x2 x0 x6 x4
102
}
103

  
104
static inline void mmxext_row (int16_t * table, int32_t * rounder)
105
{
106
    movq_m2r (*(table+8), mm1);		// mm1 = -C5 -C1 C3 C1
107
    pmaddwd_r2r (mm2, mm4);		// mm4 = C4*x0+C6*x2 C4*x4+C6*x6
108

  
109
    pmaddwd_m2r (*(table+16), mm0);	// mm0 = C4*x4-C6*x6 C4*x0-C6*x2
110
    pshufw_r2r (mm6, mm6, 0x4e);	// mm6 = x3 x1 x7 x5
111

  
112
    movq_m2r (*(table+12), mm7);	// mm7 = -C7 C3 C7 C5
113
    pmaddwd_r2r (mm5, mm1);		// mm1 = -C1*x5-C5*x7 C1*x1+C3*x3
114

  
115
    paddd_m2r (*rounder, mm3);		// mm3 += rounder
116
    pmaddwd_r2r (mm6, mm7);		// mm7 = C3*x1-C7*x3 C5*x5+C7*x7
117

  
118
    pmaddwd_m2r (*(table+20), mm2);	// mm2 = C4*x0-C2*x2 -C4*x4+C2*x6
119
    paddd_r2r (mm4, mm3);		// mm3 = a1 a0 + rounder
120

  
121
    pmaddwd_m2r (*(table+24), mm5);	// mm5 = C3*x5-C1*x7 C5*x1-C1*x3
122
    movq_r2r (mm3, mm4);		// mm4 = a1 a0 + rounder
123

  
124
    pmaddwd_m2r (*(table+28), mm6);	// mm6 = C7*x1-C5*x3 C7*x5+C3*x7
125
    paddd_r2r (mm7, mm1);		// mm1 = b1 b0
126

  
127
    paddd_m2r (*rounder, mm0);		// mm0 += rounder
128
    psubd_r2r (mm1, mm3);		// mm3 = a1-b1 a0-b0 + rounder
129

  
130
    psrad_i2r (ROW_SHIFT, mm3);		// mm3 = y6 y7
131
    paddd_r2r (mm4, mm1);		// mm1 = a1+b1 a0+b0 + rounder
132

  
133
    paddd_r2r (mm2, mm0);		// mm0 = a3 a2 + rounder
134
    psrad_i2r (ROW_SHIFT, mm1);		// mm1 = y1 y0
135

  
136
    paddd_r2r (mm6, mm5);		// mm5 = b3 b2
137
    movq_r2r (mm0, mm4);		// mm4 = a3 a2 + rounder
138

  
139
    paddd_r2r (mm5, mm0);		// mm0 = a3+b3 a2+b2 + rounder
140
    psubd_r2r (mm5, mm4);		// mm4 = a3-b3 a2-b2 + rounder
141
}
142

  
143
static inline void mmxext_row_tail (int16_t * row, int store)
144
{
145
    psrad_i2r (ROW_SHIFT, mm0);		// mm0 = y3 y2
146

  
147
    psrad_i2r (ROW_SHIFT, mm4);		// mm4 = y4 y5
148

  
149
    packssdw_r2r (mm0, mm1);		// mm1 = y3 y2 y1 y0
150

  
151
    packssdw_r2r (mm3, mm4);		// mm4 = y6 y7 y4 y5
152

  
153
    movq_r2m (mm1, *(row+store));	// save y3 y2 y1 y0
154
    pshufw_r2r (mm4, mm4, 0xb1);	// mm4 = y7 y6 y5 y4
155

  
156
    /* slot */
157

  
158
    movq_r2m (mm4, *(row+store+4));	// save y7 y6 y5 y4
159
}
160

  
161
static inline void mmxext_row_mid (int16_t * row, int store,
162
				   int offset, int16_t * table)
163
{
164
    movq_m2r (*(row+offset), mm2);	// mm2 = x6 x4 x2 x0
165
    psrad_i2r (ROW_SHIFT, mm0);		// mm0 = y3 y2
166

  
167
    movq_m2r (*(row+offset+4), mm5);	// mm5 = x7 x5 x3 x1
168
    psrad_i2r (ROW_SHIFT, mm4);		// mm4 = y4 y5
169

  
170
    packssdw_r2r (mm0, mm1);		// mm1 = y3 y2 y1 y0
171
    movq_r2r (mm5, mm6);		// mm6 = x7 x5 x3 x1
172

  
173
    packssdw_r2r (mm3, mm4);		// mm4 = y6 y7 y4 y5
174
    movq_r2r (mm2, mm0);		// mm0 = x6 x4 x2 x0
175

  
176
    movq_r2m (mm1, *(row+store));	// save y3 y2 y1 y0
177
    pshufw_r2r (mm4, mm4, 0xb1);	// mm4 = y7 y6 y5 y4
178

  
179
    movq_m2r (*table, mm3);		// mm3 = -C2 -C4 C2 C4
180
    movq_r2m (mm4, *(row+store+4));	// save y7 y6 y5 y4
181

  
182
    pmaddwd_r2r (mm0, mm3);		// mm3 = -C4*x4-C2*x6 C4*x0+C2*x2
183

  
184
    movq_m2r (*(table+4), mm4);		// mm4 = C6 C4 C6 C4
185
    pshufw_r2r (mm2, mm2, 0x4e);	// mm2 = x2 x0 x6 x4
186
}
187

  
188

  
189
/* MMX row IDCT */
190

  
191
#define mmx_table(c1,c2,c3,c4,c5,c6,c7)	{  c4,  c2,  c4,  c6,	\
192
					   c4,  c6, -c4, -c2,	\
193
					   c1,  c3,  c3, -c7,	\
194
					   c5,  c7, -c1, -c5,	\
195
					   c4, -c6,  c4, -c2,	\
196
					  -c4,  c2,  c4, -c6,	\
197
					   c5, -c1,  c7, -c5,	\
198
					   c7,  c3,  c3, -c1 }
199

  
200
static inline void mmx_row_head (int16_t * row, int offset, int16_t * table)
201
{
202
    movq_m2r (*(row+offset), mm2);	// mm2 = x6 x4 x2 x0
203

  
204
    movq_m2r (*(row+offset+4), mm5);	// mm5 = x7 x5 x3 x1
205
    movq_r2r (mm2, mm0);		// mm0 = x6 x4 x2 x0
206

  
207
    movq_m2r (*table, mm3);		// mm3 = C6 C4 C2 C4
208
    movq_r2r (mm5, mm6);		// mm6 = x7 x5 x3 x1
209

  
210
    punpckldq_r2r (mm0, mm0);		// mm0 = x2 x0 x2 x0
211

  
212
    movq_m2r (*(table+4), mm4);		// mm4 = -C2 -C4 C6 C4
213
    pmaddwd_r2r (mm0, mm3);		// mm3 = C4*x0+C6*x2 C4*x0+C2*x2
214

  
215
    movq_m2r (*(table+8), mm1);		// mm1 = -C7 C3 C3 C1
216
    punpckhdq_r2r (mm2, mm2);		// mm2 = x6 x4 x6 x4
217
}
218

  
219
static inline void mmx_row (int16_t * table, int32_t * rounder)
220
{
221
    pmaddwd_r2r (mm2, mm4);		// mm4 = -C4*x4-C2*x6 C4*x4+C6*x6
222
    punpckldq_r2r (mm5, mm5);		// mm5 = x3 x1 x3 x1
223

  
224
    pmaddwd_m2r (*(table+16), mm0);	// mm0 = C4*x0-C2*x2 C4*x0-C6*x2
225
    punpckhdq_r2r (mm6, mm6);		// mm6 = x7 x5 x7 x5
226

  
227
    movq_m2r (*(table+12), mm7);	// mm7 = -C5 -C1 C7 C5
228
    pmaddwd_r2r (mm5, mm1);		// mm1 = C3*x1-C7*x3 C1*x1+C3*x3
229

  
230
    paddd_m2r (*rounder, mm3);		// mm3 += rounder
231
    pmaddwd_r2r (mm6, mm7);		// mm7 = -C1*x5-C5*x7 C5*x5+C7*x7
232

  
233
    pmaddwd_m2r (*(table+20), mm2);	// mm2 = C4*x4-C6*x6 -C4*x4+C2*x6
234
    paddd_r2r (mm4, mm3);		// mm3 = a1 a0 + rounder
235

  
236
    pmaddwd_m2r (*(table+24), mm5);	// mm5 = C7*x1-C5*x3 C5*x1-C1*x3
237
    movq_r2r (mm3, mm4);		// mm4 = a1 a0 + rounder
238

  
239
    pmaddwd_m2r (*(table+28), mm6);	// mm6 = C3*x5-C1*x7 C7*x5+C3*x7
240
    paddd_r2r (mm7, mm1);		// mm1 = b1 b0
241

  
242
    paddd_m2r (*rounder, mm0);		// mm0 += rounder
243
    psubd_r2r (mm1, mm3);		// mm3 = a1-b1 a0-b0 + rounder
244

  
245
    psrad_i2r (ROW_SHIFT, mm3);		// mm3 = y6 y7
246
    paddd_r2r (mm4, mm1);		// mm1 = a1+b1 a0+b0 + rounder
247

  
248
    paddd_r2r (mm2, mm0);		// mm0 = a3 a2 + rounder
249
    psrad_i2r (ROW_SHIFT, mm1);		// mm1 = y1 y0
250

  
251
    paddd_r2r (mm6, mm5);		// mm5 = b3 b2
252
    movq_r2r (mm0, mm7);		// mm7 = a3 a2 + rounder
253

  
254
    paddd_r2r (mm5, mm0);		// mm0 = a3+b3 a2+b2 + rounder
255
    psubd_r2r (mm5, mm7);		// mm7 = a3-b3 a2-b2 + rounder
256
}
257

  
258
static inline void mmx_row_tail (int16_t * row, int store)
259
{
260
    psrad_i2r (ROW_SHIFT, mm0);		// mm0 = y3 y2
261

  
262
    psrad_i2r (ROW_SHIFT, mm7);		// mm7 = y4 y5
263

  
264
    packssdw_r2r (mm0, mm1);		// mm1 = y3 y2 y1 y0
265

  
266
    packssdw_r2r (mm3, mm7);		// mm7 = y6 y7 y4 y5
267

  
268
    movq_r2m (mm1, *(row+store));	// save y3 y2 y1 y0
269
    movq_r2r (mm7, mm4);		// mm4 = y6 y7 y4 y5
270

  
271
    pslld_i2r (16, mm7);		// mm7 = y7 0 y5 0
272

  
273
    psrld_i2r (16, mm4);		// mm4 = 0 y6 0 y4
274

  
275
    por_r2r (mm4, mm7);			// mm7 = y7 y6 y5 y4
276

  
277
    /* slot */
278

  
279
    movq_r2m (mm7, *(row+store+4));	// save y7 y6 y5 y4
280
}
281

  
282
static inline void mmx_row_mid (int16_t * row, int store,
283
				int offset, int16_t * table)
284
{
285
    movq_m2r (*(row+offset), mm2);	// mm2 = x6 x4 x2 x0
286
    psrad_i2r (ROW_SHIFT, mm0);		// mm0 = y3 y2
287

  
288
    movq_m2r (*(row+offset+4), mm5);	// mm5 = x7 x5 x3 x1
289
    psrad_i2r (ROW_SHIFT, mm7);		// mm7 = y4 y5
290

  
291
    packssdw_r2r (mm0, mm1);		// mm1 = y3 y2 y1 y0
292
    movq_r2r (mm5, mm6);		// mm6 = x7 x5 x3 x1
293

  
294
    packssdw_r2r (mm3, mm7);		// mm7 = y6 y7 y4 y5
295
    movq_r2r (mm2, mm0);		// mm0 = x6 x4 x2 x0
296

  
297
    movq_r2m (mm1, *(row+store));	// save y3 y2 y1 y0
298
    movq_r2r (mm7, mm1);		// mm1 = y6 y7 y4 y5
299

  
300
    punpckldq_r2r (mm0, mm0);		// mm0 = x2 x0 x2 x0
301
    psrld_i2r (16, mm7);		// mm7 = 0 y6 0 y4
302

  
303
    movq_m2r (*table, mm3);		// mm3 = C6 C4 C2 C4
304
    pslld_i2r (16, mm1);		// mm1 = y7 0 y5 0
305

  
306
    movq_m2r (*(table+4), mm4);		// mm4 = -C2 -C4 C6 C4
307
    por_r2r (mm1, mm7);			// mm7 = y7 y6 y5 y4
308

  
309
    movq_m2r (*(table+8), mm1);		// mm1 = -C7 C3 C3 C1
310
    punpckhdq_r2r (mm2, mm2);		// mm2 = x6 x4 x6 x4
311

  
312
    movq_r2m (mm7, *(row+store+4));	// save y7 y6 y5 y4
313
    pmaddwd_r2r (mm0, mm3);		// mm3 = C4*x0+C6*x2 C4*x0+C2*x2
314
}
315

  
316

  
317
#if 0
318
// C column IDCT - its just here to document the MMXEXT and MMX versions
319
static inline void idct_col (int16_t * col, int offset)
320
{
321
/* multiplication - as implemented on mmx */
322
#define F(c,x) (((c) * (x)) >> 16)
323

  
324
/* saturation - it helps us handle torture test cases */
325
#define S(x) (((x)>32767) ? 32767 : ((x)<-32768) ? -32768 : (x))
326

  
327
    int16_t x0, x1, x2, x3, x4, x5, x6, x7;
328
    int16_t y0, y1, y2, y3, y4, y5, y6, y7;
329
    int16_t a0, a1, a2, a3, b0, b1, b2, b3;
330
    int16_t u04, v04, u26, v26, u17, v17, u35, v35, u12, v12;
331

  
332
    col += offset;
333

  
334
    x0 = col[0*8];
335
    x1 = col[1*8];
336
    x2 = col[2*8];
337
    x3 = col[3*8];
338
    x4 = col[4*8];
339
    x5 = col[5*8];
340
    x6 = col[6*8];
341
    x7 = col[7*8];
342

  
343
    u04 = S (x0 + x4);
344
    v04 = S (x0 - x4);
345
    u26 = S (F (T2, x6) + x2);
346
    v26 = S (F (T2, x2) - x6);
347

  
348
    a0 = S (u04 + u26);
349
    a1 = S (v04 + v26);
350
    a2 = S (v04 - v26);
351
    a3 = S (u04 - u26);
352

  
353
    u17 = S (F (T1, x7) + x1);
354
    v17 = S (F (T1, x1) - x7);
355
    u35 = S (F (T3, x5) + x3);
356
    v35 = S (F (T3, x3) - x5);
357

  
358
    b0 = S (u17 + u35);
359
    b3 = S (v17 - v35);
360
    u12 = S (u17 - u35);
361
    v12 = S (v17 + v35);
362
    u12 = S (2 * F (C4, u12));
363
    v12 = S (2 * F (C4, v12));
364
    b1 = S (u12 + v12);
365
    b2 = S (u12 - v12);
366

  
367
    y0 = S (a0 + b0) >> COL_SHIFT;
368
    y1 = S (a1 + b1) >> COL_SHIFT;
369
    y2 = S (a2 + b2) >> COL_SHIFT;
370
    y3 = S (a3 + b3) >> COL_SHIFT;
371

  
372
    y4 = S (a3 - b3) >> COL_SHIFT;
373
    y5 = S (a2 - b2) >> COL_SHIFT;
374
    y6 = S (a1 - b1) >> COL_SHIFT;
375
    y7 = S (a0 - b0) >> COL_SHIFT;
376

  
377
    col[0*8] = y0;
378
    col[1*8] = y1;
379
    col[2*8] = y2;
380
    col[3*8] = y3;
381
    col[4*8] = y4;
382
    col[5*8] = y5;
383
    col[6*8] = y6;
384
    col[7*8] = y7;
385
}
386
#endif
387

  
388

  
389
// MMX column IDCT
390
static inline void idct_col (int16_t * col, int offset)
391
{
392
#define T1 13036
393
#define T2 27146
394
#define T3 43790
395
#define C4 23170
396

  
397
    static short _T1[] ATTR_ALIGN(8) = {T1,T1,T1,T1};
398
    static short _T2[] ATTR_ALIGN(8) = {T2,T2,T2,T2};
399
    static short _T3[] ATTR_ALIGN(8) = {T3,T3,T3,T3};
400
    static short _C4[] ATTR_ALIGN(8) = {C4,C4,C4,C4};
401

  
402
    /* column code adapted from peter gubanov */
403
    /* http://www.elecard.com/peter/idct.shtml */
404

  
405
    movq_m2r (*_T1, mm0);		// mm0 = T1
406

  
407
    movq_m2r (*(col+offset+1*8), mm1);	// mm1 = x1
408
    movq_r2r (mm0, mm2);		// mm2 = T1
409

  
410
    movq_m2r (*(col+offset+7*8), mm4);	// mm4 = x7
411
    pmulhw_r2r (mm1, mm0);		// mm0 = T1*x1
412

  
413
    movq_m2r (*_T3, mm5);		// mm5 = T3
414
    pmulhw_r2r (mm4, mm2);		// mm2 = T1*x7
415

  
416
    movq_m2r (*(col+offset+5*8), mm6);	// mm6 = x5
417
    movq_r2r (mm5, mm7);		// mm7 = T3-1
418

  
419
    movq_m2r (*(col+offset+3*8), mm3);	// mm3 = x3
420
    psubsw_r2r (mm4, mm0);		// mm0 = v17
421

  
422
    movq_m2r (*_T2, mm4);		// mm4 = T2
423
    pmulhw_r2r (mm3, mm5);		// mm5 = (T3-1)*x3
424

  
425
    paddsw_r2r (mm2, mm1);		// mm1 = u17
426
    pmulhw_r2r (mm6, mm7);		// mm7 = (T3-1)*x5
427

  
428
    /* slot */
429

  
430
    movq_r2r (mm4, mm2);		// mm2 = T2
431
    paddsw_r2r (mm3, mm5);		// mm5 = T3*x3
432

  
433
    pmulhw_m2r (*(col+offset+2*8), mm4);// mm4 = T2*x2
434
    paddsw_r2r (mm6, mm7);		// mm7 = T3*x5
435

  
436
    psubsw_r2r (mm6, mm5);		// mm5 = v35
437
    paddsw_r2r (mm3, mm7);		// mm7 = u35
438

  
439
    movq_m2r (*(col+offset+6*8), mm3);	// mm3 = x6
440
    movq_r2r (mm0, mm6);		// mm6 = v17
441

  
442
    pmulhw_r2r (mm3, mm2);		// mm2 = T2*x6
443
    psubsw_r2r (mm5, mm0);		// mm0 = b3
444

  
445
    psubsw_r2r (mm3, mm4);		// mm4 = v26
446
    paddsw_r2r (mm6, mm5);		// mm5 = v12
447

  
448
    movq_r2m (mm0, *(col+offset+3*8));	// save b3 in scratch0
449
    movq_r2r (mm1, mm6);		// mm6 = u17
450

  
451
    paddsw_m2r (*(col+offset+2*8), mm2);// mm2 = u26
452
    paddsw_r2r (mm7, mm6);		// mm6 = b0
453

  
454
    psubsw_r2r (mm7, mm1);		// mm1 = u12
455
    movq_r2r (mm1, mm7);		// mm7 = u12
456

  
457
    movq_m2r (*(col+offset+0*8), mm3);	// mm3 = x0
458
    paddsw_r2r (mm5, mm1);		// mm1 = u12+v12
459

  
460
    movq_m2r (*_C4, mm0);		// mm0 = C4/2
461
    psubsw_r2r (mm5, mm7);		// mm7 = u12-v12
462

  
463
    movq_r2m (mm6, *(col+offset+5*8));	// save b0 in scratch1
464
    pmulhw_r2r (mm0, mm1);		// mm1 = b1/2
465

  
466
    movq_r2r (mm4, mm6);		// mm6 = v26
467
    pmulhw_r2r (mm0, mm7);		// mm7 = b2/2
468

  
469
    movq_m2r (*(col+offset+4*8), mm5);	// mm5 = x4
470
    movq_r2r (mm3, mm0);		// mm0 = x0
471

  
472
    psubsw_r2r (mm5, mm3);		// mm3 = v04
473
    paddsw_r2r (mm5, mm0);		// mm0 = u04
474

  
475
    paddsw_r2r (mm3, mm4);		// mm4 = a1
476
    movq_r2r (mm0, mm5);		// mm5 = u04
477

  
478
    psubsw_r2r (mm6, mm3);		// mm3 = a2
479
    paddsw_r2r (mm2, mm5);		// mm5 = a0
480

  
481
    paddsw_r2r (mm1, mm1);		// mm1 = b1
482
    psubsw_r2r (mm2, mm0);		// mm0 = a3
483

  
484
    paddsw_r2r (mm7, mm7);		// mm7 = b2
485
    movq_r2r (mm3, mm2);		// mm2 = a2
486

  
487
    movq_r2r (mm4, mm6);		// mm6 = a1
488
    paddsw_r2r (mm7, mm3);		// mm3 = a2+b2
489

  
490
    psraw_i2r (COL_SHIFT, mm3);		// mm3 = y2
491
    paddsw_r2r (mm1, mm4);		// mm4 = a1+b1
492

  
493
    psraw_i2r (COL_SHIFT, mm4);		// mm4 = y1
494
    psubsw_r2r (mm1, mm6);		// mm6 = a1-b1
495

  
496
    movq_m2r (*(col+offset+5*8), mm1);	// mm1 = b0
497
    psubsw_r2r (mm7, mm2);		// mm2 = a2-b2
498

  
499
    psraw_i2r (COL_SHIFT, mm6);		// mm6 = y6
500
    movq_r2r (mm5, mm7);		// mm7 = a0
501

  
502
    movq_r2m (mm4, *(col+offset+1*8));	// save y1
503
    psraw_i2r (COL_SHIFT, mm2);		// mm2 = y5
504

  
505
    movq_r2m (mm3, *(col+offset+2*8));	// save y2
506
    paddsw_r2r (mm1, mm5);		// mm5 = a0+b0
507

  
508
    movq_m2r (*(col+offset+3*8), mm4);	// mm4 = b3
509
    psubsw_r2r (mm1, mm7);		// mm7 = a0-b0
510

  
511
    psraw_i2r (COL_SHIFT, mm5);		// mm5 = y0
512
    movq_r2r (mm0, mm3);		// mm3 = a3
513

  
514
    movq_r2m (mm2, *(col+offset+5*8));	// save y5
515
    psubsw_r2r (mm4, mm3);		// mm3 = a3-b3
516

  
517
    psraw_i2r (COL_SHIFT, mm7);		// mm7 = y7
518
    paddsw_r2r (mm0, mm4);		// mm4 = a3+b3
519

  
520
    movq_r2m (mm5, *(col+offset+0*8));	// save y0
521
    psraw_i2r (COL_SHIFT, mm3);		// mm3 = y4
522

  
523
    movq_r2m (mm6, *(col+offset+6*8));	// save y6
524
    psraw_i2r (COL_SHIFT, mm4);		// mm4 = y3
525

  
526
    movq_r2m (mm7, *(col+offset+7*8));	// save y7
527

  
528
    movq_r2m (mm3, *(col+offset+4*8));	// save y4
529

  
530
    movq_r2m (mm4, *(col+offset+3*8));	// save y3
531
}
532

  
533

  
534
static int32_t rounder0[] ATTR_ALIGN(8) =
535
    rounder ((1 << (COL_SHIFT - 1)) - 0.5);
536
static int32_t rounder4[] ATTR_ALIGN(8) = rounder (0);
537
static int32_t rounder1[] ATTR_ALIGN(8) =
538
    rounder (1.25683487303);	/* C1*(C1/C4+C1+C7)/2 */
539
static int32_t rounder7[] ATTR_ALIGN(8) =
540
    rounder (-0.25);		/* C1*(C7/C4+C7-C1)/2 */
541
static int32_t rounder2[] ATTR_ALIGN(8) =
542
    rounder (0.60355339059);	/* C2 * (C6+C2)/2 */
543
static int32_t rounder6[] ATTR_ALIGN(8) =
544
    rounder (-0.25);		/* C2 * (C6-C2)/2 */
545
static int32_t rounder3[] ATTR_ALIGN(8) =
546
    rounder (0.087788325588);	/* C3*(-C3/C4+C3+C5)/2 */
547
static int32_t rounder5[] ATTR_ALIGN(8) =
548
    rounder (-0.441341716183);	/* C3*(-C5/C4+C5-C3)/2 */
549

  
550

  
551
#define declare_idct(idct,table,idct_row_head,idct_row,idct_row_tail,idct_row_mid)	\
552
void idct (int16_t * block)					\
553
{									\
554
    static int16_t table04[] ATTR_ALIGN(16) =				\
555
	table (22725, 21407, 19266, 16384, 12873,  8867, 4520);		\
556
    static int16_t table17[] ATTR_ALIGN(16) =				\
557
	table (31521, 29692, 26722, 22725, 17855, 12299, 6270);		\
558
    static int16_t table26[] ATTR_ALIGN(16) =				\
559
	table (29692, 27969, 25172, 21407, 16819, 11585, 5906);		\
560
    static int16_t table35[] ATTR_ALIGN(16) =				\
561
	table (26722, 25172, 22654, 19266, 15137, 10426, 5315);		\
562
									\
563
    idct_row_head (block, 0*8, table04);				\
564
    idct_row (table04, rounder0);					\
565
    idct_row_mid (block, 0*8, 4*8, table04);				\
566
    idct_row (table04, rounder4);					\
567
    idct_row_mid (block, 4*8, 1*8, table17);				\
568
    idct_row (table17, rounder1);					\
569
    idct_row_mid (block, 1*8, 7*8, table17);				\
570
    idct_row (table17, rounder7);					\
571
    idct_row_mid (block, 7*8, 2*8, table26);				\
572
    idct_row (table26, rounder2);					\
573
    idct_row_mid (block, 2*8, 6*8, table26);				\
574
    idct_row (table26, rounder6);					\
575
    idct_row_mid (block, 6*8, 3*8, table35);				\
576
    idct_row (table35, rounder3);					\
577
    idct_row_mid (block, 3*8, 5*8, table35);				\
578
    idct_row (table35, rounder5);					\
579
    idct_row_tail (block, 5*8);						\
580
									\
581
    idct_col (block, 0);						\
582
    idct_col (block, 4);						\
583
}
584

  
585

  
586
declare_idct (ff_mmxext_idct, mmxext_table,
587
	      mmxext_row_head, mmxext_row, mmxext_row_tail, mmxext_row_mid)
588

  
589
declare_idct (ff_mmx_idct, mmx_table,
590
	      mmx_row_head, mmx_row, mmx_row_tail, mmx_row_mid)

Also available in: Unified diff