MMX for H.264 iDCT (adapted from x264)
[libav.git] / 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 "../common.h"
26 #include "../dsputil.h"
27
28 #include "mmx.h"
29
30 #define ATTR_ALIGN(align) __attribute__ ((__aligned__ (align)))
31
32 #define ROW_SHIFT 11
33 #define COL_SHIFT 6
34
35 #define round(bias) ((int)(((bias)+0.5) * (1<<ROW_SHIFT)))
36 #define rounder(bias) {round (bias), round (bias)}
37
38 #if 0
39 /* C row IDCT - its just here to document the MMXEXT and MMX versions */
40 static inline void idct_row (int16_t * row, int offset,
41 int16_t * table, int32_t * rounder)
42 {
43 int C1, C2, C3, C4, C5, C6, C7;
44 int a0, a1, a2, a3, b0, b1, b2, b3;
45
46 row += offset;
47
48 C1 = table[1];
49 C2 = table[2];
50 C3 = table[3];
51 C4 = table[4];
52 C5 = table[5];
53 C6 = table[6];
54 C7 = table[7];
55
56 a0 = C4*row[0] + C2*row[2] + C4*row[4] + C6*row[6] + *rounder;
57 a1 = C4*row[0] + C6*row[2] - C4*row[4] - C2*row[6] + *rounder;
58 a2 = C4*row[0] - C6*row[2] - C4*row[4] + C2*row[6] + *rounder;
59 a3 = C4*row[0] - C2*row[2] + C4*row[4] - C6*row[6] + *rounder;
60
61 b0 = C1*row[1] + C3*row[3] + C5*row[5] + C7*row[7];
62 b1 = C3*row[1] - C7*row[3] - C1*row[5] - C5*row[7];
63 b2 = C5*row[1] - C1*row[3] + C7*row[5] + C3*row[7];
64 b3 = C7*row[1] - C5*row[3] + C3*row[5] - C1*row[7];
65
66 row[0] = (a0 + b0) >> ROW_SHIFT;
67 row[1] = (a1 + b1) >> ROW_SHIFT;
68 row[2] = (a2 + b2) >> ROW_SHIFT;
69 row[3] = (a3 + b3) >> ROW_SHIFT;
70 row[4] = (a3 - b3) >> ROW_SHIFT;
71 row[5] = (a2 - b2) >> ROW_SHIFT;
72 row[6] = (a1 - b1) >> ROW_SHIFT;
73 row[7] = (a0 - b0) >> ROW_SHIFT;
74 }
75 #endif
76
77
78 /* MMXEXT row IDCT */
79
80 #define mmxext_table(c1,c2,c3,c4,c5,c6,c7) { c4, c2, -c4, -c2, \
81 c4, c6, c4, c6, \
82 c1, c3, -c1, -c5, \
83 c5, c7, c3, -c7, \
84 c4, -c6, c4, -c6, \
85 -c4, c2, c4, -c2, \
86 c5, -c1, c3, -c1, \
87 c7, c3, c7, -c5 }
88
89 static inline void mmxext_row_head (int16_t * row, int offset, int16_t * table)
90 {
91 movq_m2r (*(row+offset), mm2); // mm2 = x6 x4 x2 x0
92
93 movq_m2r (*(row+offset+4), mm5); // mm5 = x7 x5 x3 x1
94 movq_r2r (mm2, mm0); // mm0 = x6 x4 x2 x0
95
96 movq_m2r (*table, mm3); // mm3 = -C2 -C4 C2 C4
97 movq_r2r (mm5, mm6); // mm6 = x7 x5 x3 x1
98
99 movq_m2r (*(table+4), mm4); // mm4 = C6 C4 C6 C4
100 pmaddwd_r2r (mm0, mm3); // mm3 = -C4*x4-C2*x6 C4*x0+C2*x2
101
102 pshufw_r2r (mm2, mm2, 0x4e); // mm2 = x2 x0 x6 x4
103 }
104
105 static inline void mmxext_row (int16_t * table, int32_t * rounder)
106 {
107 movq_m2r (*(table+8), mm1); // mm1 = -C5 -C1 C3 C1
108 pmaddwd_r2r (mm2, mm4); // mm4 = C4*x0+C6*x2 C4*x4+C6*x6
109
110 pmaddwd_m2r (*(table+16), mm0); // mm0 = C4*x4-C6*x6 C4*x0-C6*x2
111 pshufw_r2r (mm6, mm6, 0x4e); // mm6 = x3 x1 x7 x5
112
113 movq_m2r (*(table+12), mm7); // mm7 = -C7 C3 C7 C5
114 pmaddwd_r2r (mm5, mm1); // mm1 = -C1*x5-C5*x7 C1*x1+C3*x3
115
116 paddd_m2r (*rounder, mm3); // mm3 += rounder
117 pmaddwd_r2r (mm6, mm7); // mm7 = C3*x1-C7*x3 C5*x5+C7*x7
118
119 pmaddwd_m2r (*(table+20), mm2); // mm2 = C4*x0-C2*x2 -C4*x4+C2*x6
120 paddd_r2r (mm4, mm3); // mm3 = a1 a0 + rounder
121
122 pmaddwd_m2r (*(table+24), mm5); // mm5 = C3*x5-C1*x7 C5*x1-C1*x3
123 movq_r2r (mm3, mm4); // mm4 = a1 a0 + rounder
124
125 pmaddwd_m2r (*(table+28), mm6); // mm6 = C7*x1-C5*x3 C7*x5+C3*x7
126 paddd_r2r (mm7, mm1); // mm1 = b1 b0
127
128 paddd_m2r (*rounder, mm0); // mm0 += rounder
129 psubd_r2r (mm1, mm3); // mm3 = a1-b1 a0-b0 + rounder
130
131 psrad_i2r (ROW_SHIFT, mm3); // mm3 = y6 y7
132 paddd_r2r (mm4, mm1); // mm1 = a1+b1 a0+b0 + rounder
133
134 paddd_r2r (mm2, mm0); // mm0 = a3 a2 + rounder
135 psrad_i2r (ROW_SHIFT, mm1); // mm1 = y1 y0
136
137 paddd_r2r (mm6, mm5); // mm5 = b3 b2
138 movq_r2r (mm0, mm4); // mm4 = a3 a2 + rounder
139
140 paddd_r2r (mm5, mm0); // mm0 = a3+b3 a2+b2 + rounder
141 psubd_r2r (mm5, mm4); // mm4 = a3-b3 a2-b2 + rounder
142 }
143
144 static inline void mmxext_row_tail (int16_t * row, int store)
145 {
146 psrad_i2r (ROW_SHIFT, mm0); // mm0 = y3 y2
147
148 psrad_i2r (ROW_SHIFT, mm4); // mm4 = y4 y5
149
150 packssdw_r2r (mm0, mm1); // mm1 = y3 y2 y1 y0
151
152 packssdw_r2r (mm3, mm4); // mm4 = y6 y7 y4 y5
153
154 movq_r2m (mm1, *(row+store)); // save y3 y2 y1 y0
155 pshufw_r2r (mm4, mm4, 0xb1); // mm4 = y7 y6 y5 y4
156
157 /* slot */
158
159 movq_r2m (mm4, *(row+store+4)); // save y7 y6 y5 y4
160 }
161
162 static inline void mmxext_row_mid (int16_t * row, int store,
163 int offset, int16_t * table)
164 {
165 movq_m2r (*(row+offset), mm2); // mm2 = x6 x4 x2 x0
166 psrad_i2r (ROW_SHIFT, mm0); // mm0 = y3 y2
167
168 movq_m2r (*(row+offset+4), mm5); // mm5 = x7 x5 x3 x1
169 psrad_i2r (ROW_SHIFT, mm4); // mm4 = y4 y5
170
171 packssdw_r2r (mm0, mm1); // mm1 = y3 y2 y1 y0
172 movq_r2r (mm5, mm6); // mm6 = x7 x5 x3 x1
173
174 packssdw_r2r (mm3, mm4); // mm4 = y6 y7 y4 y5
175 movq_r2r (mm2, mm0); // mm0 = x6 x4 x2 x0
176
177 movq_r2m (mm1, *(row+store)); // save y3 y2 y1 y0
178 pshufw_r2r (mm4, mm4, 0xb1); // mm4 = y7 y6 y5 y4
179
180 movq_m2r (*table, mm3); // mm3 = -C2 -C4 C2 C4
181 movq_r2m (mm4, *(row+store+4)); // save y7 y6 y5 y4
182
183 pmaddwd_r2r (mm0, mm3); // mm3 = -C4*x4-C2*x6 C4*x0+C2*x2
184
185 movq_m2r (*(table+4), mm4); // mm4 = C6 C4 C6 C4
186 pshufw_r2r (mm2, mm2, 0x4e); // mm2 = x2 x0 x6 x4
187 }
188
189
190 /* MMX row IDCT */
191
192 #define mmx_table(c1,c2,c3,c4,c5,c6,c7) { c4, c2, c4, c6, \
193 c4, c6, -c4, -c2, \
194 c1, c3, c3, -c7, \
195 c5, c7, -c1, -c5, \
196 c4, -c6, c4, -c2, \
197 -c4, c2, c4, -c6, \
198 c5, -c1, c7, -c5, \
199 c7, c3, c3, -c1 }
200
201 static inline void mmx_row_head (int16_t * row, int offset, int16_t * table)
202 {
203 movq_m2r (*(row+offset), mm2); // mm2 = x6 x4 x2 x0
204
205 movq_m2r (*(row+offset+4), mm5); // mm5 = x7 x5 x3 x1
206 movq_r2r (mm2, mm0); // mm0 = x6 x4 x2 x0
207
208 movq_m2r (*table, mm3); // mm3 = C6 C4 C2 C4
209 movq_r2r (mm5, mm6); // mm6 = x7 x5 x3 x1
210
211 punpckldq_r2r (mm0, mm0); // mm0 = x2 x0 x2 x0
212
213 movq_m2r (*(table+4), mm4); // mm4 = -C2 -C4 C6 C4
214 pmaddwd_r2r (mm0, mm3); // mm3 = C4*x0+C6*x2 C4*x0+C2*x2
215
216 movq_m2r (*(table+8), mm1); // mm1 = -C7 C3 C3 C1
217 punpckhdq_r2r (mm2, mm2); // mm2 = x6 x4 x6 x4
218 }
219
220 static inline void mmx_row (int16_t * table, int32_t * rounder)
221 {
222 pmaddwd_r2r (mm2, mm4); // mm4 = -C4*x4-C2*x6 C4*x4+C6*x6
223 punpckldq_r2r (mm5, mm5); // mm5 = x3 x1 x3 x1
224
225 pmaddwd_m2r (*(table+16), mm0); // mm0 = C4*x0-C2*x2 C4*x0-C6*x2
226 punpckhdq_r2r (mm6, mm6); // mm6 = x7 x5 x7 x5
227
228 movq_m2r (*(table+12), mm7); // mm7 = -C5 -C1 C7 C5
229 pmaddwd_r2r (mm5, mm1); // mm1 = C3*x1-C7*x3 C1*x1+C3*x3
230
231 paddd_m2r (*rounder, mm3); // mm3 += rounder
232 pmaddwd_r2r (mm6, mm7); // mm7 = -C1*x5-C5*x7 C5*x5+C7*x7
233
234 pmaddwd_m2r (*(table+20), mm2); // mm2 = C4*x4-C6*x6 -C4*x4+C2*x6
235 paddd_r2r (mm4, mm3); // mm3 = a1 a0 + rounder
236
237 pmaddwd_m2r (*(table+24), mm5); // mm5 = C7*x1-C5*x3 C5*x1-C1*x3
238 movq_r2r (mm3, mm4); // mm4 = a1 a0 + rounder
239
240 pmaddwd_m2r (*(table+28), mm6); // mm6 = C3*x5-C1*x7 C7*x5+C3*x7
241 paddd_r2r (mm7, mm1); // mm1 = b1 b0
242
243 paddd_m2r (*rounder, mm0); // mm0 += rounder
244 psubd_r2r (mm1, mm3); // mm3 = a1-b1 a0-b0 + rounder
245
246 psrad_i2r (ROW_SHIFT, mm3); // mm3 = y6 y7
247 paddd_r2r (mm4, mm1); // mm1 = a1+b1 a0+b0 + rounder
248
249 paddd_r2r (mm2, mm0); // mm0 = a3 a2 + rounder
250 psrad_i2r (ROW_SHIFT, mm1); // mm1 = y1 y0
251
252 paddd_r2r (mm6, mm5); // mm5 = b3 b2
253 movq_r2r (mm0, mm7); // mm7 = a3 a2 + rounder
254
255 paddd_r2r (mm5, mm0); // mm0 = a3+b3 a2+b2 + rounder
256 psubd_r2r (mm5, mm7); // mm7 = a3-b3 a2-b2 + rounder
257 }
258
259 static inline void mmx_row_tail (int16_t * row, int store)
260 {
261 psrad_i2r (ROW_SHIFT, mm0); // mm0 = y3 y2
262
263 psrad_i2r (ROW_SHIFT, mm7); // mm7 = y4 y5
264
265 packssdw_r2r (mm0, mm1); // mm1 = y3 y2 y1 y0
266
267 packssdw_r2r (mm3, mm7); // mm7 = y6 y7 y4 y5
268
269 movq_r2m (mm1, *(row+store)); // save y3 y2 y1 y0
270 movq_r2r (mm7, mm4); // mm4 = y6 y7 y4 y5
271
272 pslld_i2r (16, mm7); // mm7 = y7 0 y5 0
273
274 psrld_i2r (16, mm4); // mm4 = 0 y6 0 y4
275
276 por_r2r (mm4, mm7); // mm7 = y7 y6 y5 y4
277
278 /* slot */
279
280 movq_r2m (mm7, *(row+store+4)); // save y7 y6 y5 y4
281 }
282
283 static inline void mmx_row_mid (int16_t * row, int store,
284 int offset, int16_t * table)
285 {
286 movq_m2r (*(row+offset), mm2); // mm2 = x6 x4 x2 x0
287 psrad_i2r (ROW_SHIFT, mm0); // mm0 = y3 y2
288
289 movq_m2r (*(row+offset+4), mm5); // mm5 = x7 x5 x3 x1
290 psrad_i2r (ROW_SHIFT, mm7); // mm7 = y4 y5
291
292 packssdw_r2r (mm0, mm1); // mm1 = y3 y2 y1 y0
293 movq_r2r (mm5, mm6); // mm6 = x7 x5 x3 x1
294
295 packssdw_r2r (mm3, mm7); // mm7 = y6 y7 y4 y5
296 movq_r2r (mm2, mm0); // mm0 = x6 x4 x2 x0
297
298 movq_r2m (mm1, *(row+store)); // save y3 y2 y1 y0
299 movq_r2r (mm7, mm1); // mm1 = y6 y7 y4 y5
300
301 punpckldq_r2r (mm0, mm0); // mm0 = x2 x0 x2 x0
302 psrld_i2r (16, mm7); // mm7 = 0 y6 0 y4
303
304 movq_m2r (*table, mm3); // mm3 = C6 C4 C2 C4
305 pslld_i2r (16, mm1); // mm1 = y7 0 y5 0
306
307 movq_m2r (*(table+4), mm4); // mm4 = -C2 -C4 C6 C4
308 por_r2r (mm1, mm7); // mm7 = y7 y6 y5 y4
309
310 movq_m2r (*(table+8), mm1); // mm1 = -C7 C3 C3 C1
311 punpckhdq_r2r (mm2, mm2); // mm2 = x6 x4 x6 x4
312
313 movq_r2m (mm7, *(row+store+4)); // save y7 y6 y5 y4
314 pmaddwd_r2r (mm0, mm3); // mm3 = C4*x0+C6*x2 C4*x0+C2*x2
315 }
316
317
318 #if 0
319 // C column IDCT - its just here to document the MMXEXT and MMX versions
320 static inline void idct_col (int16_t * col, int offset)
321 {
322 /* multiplication - as implemented on mmx */
323 #define F(c,x) (((c) * (x)) >> 16)
324
325 /* saturation - it helps us handle torture test cases */
326 #define S(x) (((x)>32767) ? 32767 : ((x)<-32768) ? -32768 : (x))
327
328 int16_t x0, x1, x2, x3, x4, x5, x6, x7;
329 int16_t y0, y1, y2, y3, y4, y5, y6, y7;
330 int16_t a0, a1, a2, a3, b0, b1, b2, b3;
331 int16_t u04, v04, u26, v26, u17, v17, u35, v35, u12, v12;
332
333 col += offset;
334
335 x0 = col[0*8];
336 x1 = col[1*8];
337 x2 = col[2*8];
338 x3 = col[3*8];
339 x4 = col[4*8];
340 x5 = col[5*8];
341 x6 = col[6*8];
342 x7 = col[7*8];
343
344 u04 = S (x0 + x4);
345 v04 = S (x0 - x4);
346 u26 = S (F (T2, x6) + x2);
347 v26 = S (F (T2, x2) - x6);
348
349 a0 = S (u04 + u26);
350 a1 = S (v04 + v26);
351 a2 = S (v04 - v26);
352 a3 = S (u04 - u26);
353
354 u17 = S (F (T1, x7) + x1);
355 v17 = S (F (T1, x1) - x7);
356 u35 = S (F (T3, x5) + x3);
357 v35 = S (F (T3, x3) - x5);
358
359 b0 = S (u17 + u35);
360 b3 = S (v17 - v35);
361 u12 = S (u17 - u35);
362 v12 = S (v17 + v35);
363 u12 = S (2 * F (C4, u12));
364 v12 = S (2 * F (C4, v12));
365 b1 = S (u12 + v12);
366 b2 = S (u12 - v12);
367
368 y0 = S (a0 + b0) >> COL_SHIFT;
369 y1 = S (a1 + b1) >> COL_SHIFT;
370 y2 = S (a2 + b2) >> COL_SHIFT;
371 y3 = S (a3 + b3) >> COL_SHIFT;
372
373 y4 = S (a3 - b3) >> COL_SHIFT;
374 y5 = S (a2 - b2) >> COL_SHIFT;
375 y6 = S (a1 - b1) >> COL_SHIFT;
376 y7 = S (a0 - b0) >> COL_SHIFT;
377
378 col[0*8] = y0;
379 col[1*8] = y1;
380 col[2*8] = y2;
381 col[3*8] = y3;
382 col[4*8] = y4;
383 col[5*8] = y5;
384 col[6*8] = y6;
385 col[7*8] = y7;
386 }
387 #endif
388
389
390 // MMX column IDCT
391 static inline void idct_col (int16_t * col, int offset)
392 {
393 #define T1 13036
394 #define T2 27146
395 #define T3 43790
396 #define C4 23170
397
398 static short _T1[] ATTR_ALIGN(8) = {T1,T1,T1,T1};
399 static short _T2[] ATTR_ALIGN(8) = {T2,T2,T2,T2};
400 static short _T3[] ATTR_ALIGN(8) = {T3,T3,T3,T3};
401 static short _C4[] ATTR_ALIGN(8) = {C4,C4,C4,C4};
402
403 /* column code adapted from peter gubanov */
404 /* http://www.elecard.com/peter/idct.shtml */
405
406 movq_m2r (*_T1, mm0); // mm0 = T1
407
408 movq_m2r (*(col+offset+1*8), mm1); // mm1 = x1
409 movq_r2r (mm0, mm2); // mm2 = T1
410
411 movq_m2r (*(col+offset+7*8), mm4); // mm4 = x7
412 pmulhw_r2r (mm1, mm0); // mm0 = T1*x1
413
414 movq_m2r (*_T3, mm5); // mm5 = T3
415 pmulhw_r2r (mm4, mm2); // mm2 = T1*x7
416
417 movq_m2r (*(col+offset+5*8), mm6); // mm6 = x5
418 movq_r2r (mm5, mm7); // mm7 = T3-1
419
420 movq_m2r (*(col+offset+3*8), mm3); // mm3 = x3
421 psubsw_r2r (mm4, mm0); // mm0 = v17
422
423 movq_m2r (*_T2, mm4); // mm4 = T2
424 pmulhw_r2r (mm3, mm5); // mm5 = (T3-1)*x3
425
426 paddsw_r2r (mm2, mm1); // mm1 = u17
427 pmulhw_r2r (mm6, mm7); // mm7 = (T3-1)*x5
428
429 /* slot */
430
431 movq_r2r (mm4, mm2); // mm2 = T2
432 paddsw_r2r (mm3, mm5); // mm5 = T3*x3
433
434 pmulhw_m2r (*(col+offset+2*8), mm4);// mm4 = T2*x2
435 paddsw_r2r (mm6, mm7); // mm7 = T3*x5
436
437 psubsw_r2r (mm6, mm5); // mm5 = v35
438 paddsw_r2r (mm3, mm7); // mm7 = u35
439
440 movq_m2r (*(col+offset+6*8), mm3); // mm3 = x6
441 movq_r2r (mm0, mm6); // mm6 = v17
442
443 pmulhw_r2r (mm3, mm2); // mm2 = T2*x6
444 psubsw_r2r (mm5, mm0); // mm0 = b3
445
446 psubsw_r2r (mm3, mm4); // mm4 = v26
447 paddsw_r2r (mm6, mm5); // mm5 = v12
448
449 movq_r2m (mm0, *(col+offset+3*8)); // save b3 in scratch0
450 movq_r2r (mm1, mm6); // mm6 = u17
451
452 paddsw_m2r (*(col+offset+2*8), mm2);// mm2 = u26
453 paddsw_r2r (mm7, mm6); // mm6 = b0
454
455 psubsw_r2r (mm7, mm1); // mm1 = u12
456 movq_r2r (mm1, mm7); // mm7 = u12
457
458 movq_m2r (*(col+offset+0*8), mm3); // mm3 = x0
459 paddsw_r2r (mm5, mm1); // mm1 = u12+v12
460
461 movq_m2r (*_C4, mm0); // mm0 = C4/2
462 psubsw_r2r (mm5, mm7); // mm7 = u12-v12
463
464 movq_r2m (mm6, *(col+offset+5*8)); // save b0 in scratch1
465 pmulhw_r2r (mm0, mm1); // mm1 = b1/2
466
467 movq_r2r (mm4, mm6); // mm6 = v26
468 pmulhw_r2r (mm0, mm7); // mm7 = b2/2
469
470 movq_m2r (*(col+offset+4*8), mm5); // mm5 = x4
471 movq_r2r (mm3, mm0); // mm0 = x0
472
473 psubsw_r2r (mm5, mm3); // mm3 = v04
474 paddsw_r2r (mm5, mm0); // mm0 = u04
475
476 paddsw_r2r (mm3, mm4); // mm4 = a1
477 movq_r2r (mm0, mm5); // mm5 = u04
478
479 psubsw_r2r (mm6, mm3); // mm3 = a2
480 paddsw_r2r (mm2, mm5); // mm5 = a0
481
482 paddsw_r2r (mm1, mm1); // mm1 = b1
483 psubsw_r2r (mm2, mm0); // mm0 = a3
484
485 paddsw_r2r (mm7, mm7); // mm7 = b2
486 movq_r2r (mm3, mm2); // mm2 = a2
487
488 movq_r2r (mm4, mm6); // mm6 = a1
489 paddsw_r2r (mm7, mm3); // mm3 = a2+b2
490
491 psraw_i2r (COL_SHIFT, mm3); // mm3 = y2
492 paddsw_r2r (mm1, mm4); // mm4 = a1+b1
493
494 psraw_i2r (COL_SHIFT, mm4); // mm4 = y1
495 psubsw_r2r (mm1, mm6); // mm6 = a1-b1
496
497 movq_m2r (*(col+offset+5*8), mm1); // mm1 = b0
498 psubsw_r2r (mm7, mm2); // mm2 = a2-b2
499
500 psraw_i2r (COL_SHIFT, mm6); // mm6 = y6
501 movq_r2r (mm5, mm7); // mm7 = a0
502
503 movq_r2m (mm4, *(col+offset+1*8)); // save y1
504 psraw_i2r (COL_SHIFT, mm2); // mm2 = y5
505
506 movq_r2m (mm3, *(col+offset+2*8)); // save y2
507 paddsw_r2r (mm1, mm5); // mm5 = a0+b0
508
509 movq_m2r (*(col+offset+3*8), mm4); // mm4 = b3
510 psubsw_r2r (mm1, mm7); // mm7 = a0-b0
511
512 psraw_i2r (COL_SHIFT, mm5); // mm5 = y0
513 movq_r2r (mm0, mm3); // mm3 = a3
514
515 movq_r2m (mm2, *(col+offset+5*8)); // save y5
516 psubsw_r2r (mm4, mm3); // mm3 = a3-b3
517
518 psraw_i2r (COL_SHIFT, mm7); // mm7 = y7
519 paddsw_r2r (mm0, mm4); // mm4 = a3+b3
520
521 movq_r2m (mm5, *(col+offset+0*8)); // save y0
522 psraw_i2r (COL_SHIFT, mm3); // mm3 = y4
523
524 movq_r2m (mm6, *(col+offset+6*8)); // save y6
525 psraw_i2r (COL_SHIFT, mm4); // mm4 = y3
526
527 movq_r2m (mm7, *(col+offset+7*8)); // save y7
528
529 movq_r2m (mm3, *(col+offset+4*8)); // save y4
530
531 movq_r2m (mm4, *(col+offset+3*8)); // save y3
532
533 #undef T1
534 #undef T2
535 #undef T3
536 #undef C4
537 }
538
539 static int32_t rounder0[] ATTR_ALIGN(8) =
540 rounder ((1 << (COL_SHIFT - 1)) - 0.5);
541 static int32_t rounder4[] ATTR_ALIGN(8) = rounder (0);
542 static int32_t rounder1[] ATTR_ALIGN(8) =
543 rounder (1.25683487303); /* C1*(C1/C4+C1+C7)/2 */
544 static int32_t rounder7[] ATTR_ALIGN(8) =
545 rounder (-0.25); /* C1*(C7/C4+C7-C1)/2 */
546 static int32_t rounder2[] ATTR_ALIGN(8) =
547 rounder (0.60355339059); /* C2 * (C6+C2)/2 */
548 static int32_t rounder6[] ATTR_ALIGN(8) =
549 rounder (-0.25); /* C2 * (C6-C2)/2 */
550 static int32_t rounder3[] ATTR_ALIGN(8) =
551 rounder (0.087788325588); /* C3*(-C3/C4+C3+C5)/2 */
552 static int32_t rounder5[] ATTR_ALIGN(8) =
553 rounder (-0.441341716183); /* C3*(-C5/C4+C5-C3)/2 */
554
555 #undef COL_SHIFT
556 #undef ROW_SHIFT
557
558 #define declare_idct(idct,table,idct_row_head,idct_row,idct_row_tail,idct_row_mid) \
559 void idct (int16_t * block) \
560 { \
561 static int16_t table04[] ATTR_ALIGN(16) = \
562 table (22725, 21407, 19266, 16384, 12873, 8867, 4520); \
563 static int16_t table17[] ATTR_ALIGN(16) = \
564 table (31521, 29692, 26722, 22725, 17855, 12299, 6270); \
565 static int16_t table26[] ATTR_ALIGN(16) = \
566 table (29692, 27969, 25172, 21407, 16819, 11585, 5906); \
567 static int16_t table35[] ATTR_ALIGN(16) = \
568 table (26722, 25172, 22654, 19266, 15137, 10426, 5315); \
569 \
570 idct_row_head (block, 0*8, table04); \
571 idct_row (table04, rounder0); \
572 idct_row_mid (block, 0*8, 4*8, table04); \
573 idct_row (table04, rounder4); \
574 idct_row_mid (block, 4*8, 1*8, table17); \
575 idct_row (table17, rounder1); \
576 idct_row_mid (block, 1*8, 7*8, table17); \
577 idct_row (table17, rounder7); \
578 idct_row_mid (block, 7*8, 2*8, table26); \
579 idct_row (table26, rounder2); \
580 idct_row_mid (block, 2*8, 6*8, table26); \
581 idct_row (table26, rounder6); \
582 idct_row_mid (block, 6*8, 3*8, table35); \
583 idct_row (table35, rounder3); \
584 idct_row_mid (block, 3*8, 5*8, table35); \
585 idct_row (table35, rounder5); \
586 idct_row_tail (block, 5*8); \
587 \
588 idct_col (block, 0); \
589 idct_col (block, 4); \
590 }
591
592 void ff_mmx_idct(DCTELEM *block);
593 void ff_mmxext_idct(DCTELEM *block);
594
595 declare_idct (ff_mmxext_idct, mmxext_table,
596 mmxext_row_head, mmxext_row, mmxext_row_tail, mmxext_row_mid)
597
598 declare_idct (ff_mmx_idct, mmx_table,
599 mmx_row_head, mmx_row, mmx_row_tail, mmx_row_mid)
600
601
602
603 /* in/out: mma=mma+mmb, mmb=mmb-mma */
604 #define SUMSUB_BA( a, b ) \
605 "paddw "#b", "#a" \n\t"\
606 "paddw "#b", "#b" \n\t"\
607 "psubw "#a", "#b" \n\t"
608
609 #define SUMSUB_BADC( a, b, c, d ) \
610 "paddw "#b", "#a" \n\t"\
611 "paddw "#d", "#c" \n\t"\
612 "paddw "#b", "#b" \n\t"\
613 "paddw "#d", "#d" \n\t"\
614 "psubw "#a", "#b" \n\t"\
615 "psubw "#c", "#d" \n\t"
616
617 /* in: a,b out: a,s */
618 #define SUMSUBD2_AB( a, b, t, s ) \
619 "movq "#a", "#s" \n\t"\
620 "movq "#b", "#t" \n\t"\
621 "psraw $1 , "#b" \n\t"\
622 "psraw $1 , "#s" \n\t"\
623 "paddw "#b", "#a" \n\t"\
624 "psubw "#t", "#s" \n\t"
625
626 #define IDCT4_1D( s02, s13, d02, d13, t, u ) \
627 SUMSUB_BA ( s02, d02 )\
628 SUMSUBD2_AB( s13, d13, u, t )\
629 SUMSUB_BADC( s13, s02, t, d02 )
630
631 #define SBUTTERFLY( a, b, t, n ) \
632 "movq "#a", "#t" \n\t" /* abcd */\
633 "punpckl"#n" "#b", "#a" \n\t" /* aebf */\
634 "punpckh"#n" "#b", "#t" \n\t" /* cgdh */
635
636 #define TRANSPOSE4( a, b, c, d, t ) \
637 SBUTTERFLY( a, b, t, wd ) /* a=aebf t=cgdh */\
638 SBUTTERFLY( c, d, b, wd ) /* c=imjn b=kolp */\
639 SBUTTERFLY( a, c, d, dq ) /* a=aeim d=bfjn */\
640 SBUTTERFLY( t, b, c, dq ) /* t=cgko c=dhlp */
641
642 #define STORE_DIFF_4P( p, t, pw32, z, dst ) \
643 asm volatile(\
644 "paddw "#pw32", "#p" \n\t"\
645 "psraw $6, "#p" \n\t"\
646 "movd (%0), "#t" \n\t"\
647 "punpcklbw "#z", "#t" \n\t"\
648 "paddsw "#t", "#p" \n\t"\
649 "packuswb "#z", "#p" \n\t"\
650 "movd "#p", (%0) \n\t" :: "r"(dst) )
651
652 static const uint64_t ff_pw_32 attribute_used __attribute__ ((aligned(8))) = 0x0020002000200020ULL;
653
654 void ff_h264_idct_add_mmx2(uint8_t *dst, int16_t *block, int stride)
655 {
656 /* Load dct coeffs */
657 asm volatile(
658 "movq (%0), %%mm0 \n\t"
659 "movq 8(%0), %%mm1 \n\t"
660 "movq 16(%0), %%mm2 \n\t"
661 "movq 24(%0), %%mm3 \n\t"
662 :: "r"(block) );
663
664 asm volatile(
665 /* mm1=s02+s13 mm2=s02-s13 mm4=d02+d13 mm0=d02-d13 */
666 IDCT4_1D( %%mm2, %%mm1, %%mm0, %%mm3, %%mm4, %%mm5 )
667
668 /* in: 1,4,0,2 out: 1,2,3,0 */
669 TRANSPOSE4( %%mm1, %%mm4, %%mm0, %%mm2, %%mm3 )
670
671 /* mm2=s02+s13 mm3=s02-s13 mm4=d02+d13 mm1=d02-d13 */
672 IDCT4_1D( %%mm3, %%mm2, %%mm1, %%mm0, %%mm4, %%mm5 )
673
674 /* in: 2,4,1,3 out: 2,3,0,1 */
675 TRANSPOSE4( %%mm2, %%mm4, %%mm1, %%mm3, %%mm0 )
676
677 "pxor %%mm7, %%mm7 \n\t"
678 "movq ff_pw_32, %%mm6 \n\t"
679 :: );
680
681 STORE_DIFF_4P( %%mm2, %%mm4, %%mm6, %%mm7, &dst[0*stride] );
682 STORE_DIFF_4P( %%mm3, %%mm4, %%mm6, %%mm7, &dst[1*stride] );
683 STORE_DIFF_4P( %%mm0, %%mm4, %%mm6, %%mm7, &dst[2*stride] );
684 STORE_DIFF_4P( %%mm1, %%mm4, %%mm6, %%mm7, &dst[3*stride] );
685 }