FFmpeg
 All Data Structures Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
idct_mmx.c
Go to the documentation of this file.
1 /*
2  * Copyright (C) 1999-2001 Aaron Holtzman <aholtzma@ess.engr.uvic.ca>
3  *
4  * This file is part of mpeg2dec, a free MPEG-2 video stream decoder.
5  * See http://libmpeg2.sourceforge.net/ for updates.
6  *
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.
11  *
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.
16  *
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
20  */
21 
22 #include "libavutil/common.h"
23 #include "libavcodec/dsputil.h"
24 
25 #include "libavutil/x86/asm.h"
26 #include "dsputil_mmx.h"
27 
28 #if HAVE_INLINE_ASM
29 
30 #define ROW_SHIFT 11
31 #define COL_SHIFT 6
32 
33 #define round(bias) ((int)(((bias)+0.5) * (1<<ROW_SHIFT)))
34 #define rounder(bias) {round (bias), round (bias)}
35 
36 
37 #if 0
38 /* C row IDCT - it is 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 * const row, const int offset,
89  const int16_t * const table)
90 {
91  __asm__ volatile(
92  "movq (%0), %%mm2 \n\t" /* mm2 = x6 x4 x2 x0 */
93 
94  "movq 8(%0), %%mm5 \n\t" /* mm5 = x7 x5 x3 x1 */
95  "movq %%mm2, %%mm0 \n\t" /* mm0 = x6 x4 x2 x0 */
96 
97  "movq (%1), %%mm3 \n\t" /* mm3 = -C2 -C4 C2 C4 */
98  "movq %%mm5, %%mm6 \n\t" /* mm6 = x7 x5 x3 x1 */
99 
100  "movq 8(%1), %%mm4 \n\t" /* mm4 = C6 C4 C6 C4 */
101  "pmaddwd %%mm0, %%mm3 \n\t" /* mm3 = -C4*x4-C2*x6 C4*x0+C2*x2 */
102 
103  "pshufw $0x4e, %%mm2, %%mm2 \n\t" /* mm2 = x2 x0 x6 x4 */
104  :: "r" ((row+offset)), "r" (table)
105  );
106 }
107 
108 static inline void mmxext_row (const int16_t * const table,
109  const int32_t * const rounder)
110 {
111  __asm__ volatile (
112  "movq 16(%0), %%mm1 \n\t" /* mm1 = -C5 -C1 C3 C1 */
113  "pmaddwd %%mm2, %%mm4 \n\t" /* mm4 = C4*x0+C6*x2 C4*x4+C6*x6 */
114 
115  "pmaddwd 32(%0), %%mm0 \n\t" /* mm0 = C4*x4-C6*x6 C4*x0-C6*x2 */
116  "pshufw $0x4e, %%mm6, %%mm6 \n\t" /* mm6 = x3 x1 x7 x5 */
117 
118  "movq 24(%0), %%mm7 \n\t" /* mm7 = -C7 C3 C7 C5 */
119  "pmaddwd %%mm5, %%mm1 \n\t" /* mm1= -C1*x5-C5*x7 C1*x1+C3*x3 */
120 
121  "paddd (%1), %%mm3 \n\t" /* mm3 += rounder */
122  "pmaddwd %%mm6, %%mm7 \n\t" /* mm7 = C3*x1-C7*x3 C5*x5+C7*x7 */
123 
124  "pmaddwd 40(%0), %%mm2 \n\t" /* mm2= C4*x0-C2*x2 -C4*x4+C2*x6 */
125  "paddd %%mm4, %%mm3 \n\t" /* mm3 = a1 a0 + rounder */
126 
127  "pmaddwd 48(%0), %%mm5 \n\t" /* mm5 = C3*x5-C1*x7 C5*x1-C1*x3 */
128  "movq %%mm3, %%mm4 \n\t" /* mm4 = a1 a0 + rounder */
129 
130  "pmaddwd 56(%0), %%mm6 \n\t" /* mm6 = C7*x1-C5*x3 C7*x5+C3*x7 */
131  "paddd %%mm7, %%mm1 \n\t" /* mm1 = b1 b0 */
132 
133  "paddd (%1), %%mm0 \n\t" /* mm0 += rounder */
134  "psubd %%mm1, %%mm3 \n\t" /* mm3 = a1-b1 a0-b0 + rounder */
135 
136  "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm3 \n\t" /* mm3 = y6 y7 */
137  "paddd %%mm4, %%mm1 \n\t" /* mm1 = a1+b1 a0+b0 + rounder */
138 
139  "paddd %%mm2, %%mm0 \n\t" /* mm0 = a3 a2 + rounder */
140  "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm1 \n\t" /* mm1 = y1 y0 */
141 
142  "paddd %%mm6, %%mm5 \n\t" /* mm5 = b3 b2 */
143  "movq %%mm0, %%mm4 \n\t" /* mm4 = a3 a2 + rounder */
144 
145  "paddd %%mm5, %%mm0 \n\t" /* mm0 = a3+b3 a2+b2 + rounder */
146  "psubd %%mm5, %%mm4 \n\t" /* mm4 = a3-b3 a2-b2 + rounder */
147  : : "r" (table), "r" (rounder));
148 }
149 
150 static inline void mmxext_row_tail (int16_t * const row, const int store)
151 {
152  __asm__ volatile (
153  "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t" /* mm0 = y3 y2 */
154 
155  "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm4 \n\t" /* mm4 = y4 y5 */
156 
157  "packssdw %%mm0, %%mm1 \n\t" /* mm1 = y3 y2 y1 y0 */
158 
159  "packssdw %%mm3, %%mm4 \n\t" /* mm4 = y6 y7 y4 y5 */
160 
161  "movq %%mm1, (%0) \n\t" /* save y3 y2 y1 y0 */
162  "pshufw $0xb1, %%mm4, %%mm4 \n\t" /* mm4 = y7 y6 y5 y4 */
163 
164  /* slot */
165 
166  "movq %%mm4, 8(%0) \n\t" /* save y7 y6 y5 y4 */
167  :: "r" (row+store)
168  );
169 }
170 
171 static inline void mmxext_row_mid (int16_t * const row, const int store,
172  const int offset,
173  const int16_t * const table)
174 {
175  __asm__ volatile (
176  "movq (%0,%1), %%mm2 \n\t" /* mm2 = x6 x4 x2 x0 */
177  "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t" /* mm0 = y3 y2 */
178 
179  "movq 8(%0,%1), %%mm5 \n\t" /* mm5 = x7 x5 x3 x1 */
180  "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm4 \n\t" /* mm4 = y4 y5 */
181 
182  "packssdw %%mm0, %%mm1 \n\t" /* mm1 = y3 y2 y1 y0 */
183  "movq %%mm5, %%mm6 \n\t" /* mm6 = x7 x5 x3 x1 */
184 
185  "packssdw %%mm3, %%mm4 \n\t" /* mm4 = y6 y7 y4 y5 */
186  "movq %%mm2, %%mm0 \n\t" /* mm0 = x6 x4 x2 x0 */
187 
188  "movq %%mm1, (%0,%2) \n\t" /* save y3 y2 y1 y0 */
189  "pshufw $0xb1, %%mm4, %%mm4\n\t" /* mm4 = y7 y6 y5 y4 */
190 
191  "movq (%3), %%mm3 \n\t" /* mm3 = -C2 -C4 C2 C4 */
192  "movq %%mm4, 8(%0,%2) \n\t" /* save y7 y6 y5 y4 */
193 
194  "pmaddwd %%mm0, %%mm3 \n\t" /* mm3= -C4*x4-C2*x6 C4*x0+C2*x2 */
195 
196  "movq 8(%3), %%mm4 \n\t" /* mm4 = C6 C4 C6 C4 */
197  "pshufw $0x4e, %%mm2, %%mm2\n\t" /* mm2 = x2 x0 x6 x4 */
198  :: "r" (row), "r" ((x86_reg) (2*offset)), "r" ((x86_reg) (2*store)), "r" (table)
199  );
200 }
201 
202 
203 /* MMX row IDCT */
204 
205 #define mmx_table(c1,c2,c3,c4,c5,c6,c7) { c4, c2, c4, c6, \
206  c4, c6, -c4, -c2, \
207  c1, c3, c3, -c7, \
208  c5, c7, -c1, -c5, \
209  c4, -c6, c4, -c2, \
210  -c4, c2, c4, -c6, \
211  c5, -c1, c7, -c5, \
212  c7, c3, c3, -c1 }
213 
214 static inline void mmx_row_head (int16_t * const row, const int offset,
215  const int16_t * const table)
216 {
217  __asm__ volatile (
218  "movq (%0), %%mm2 \n\t" /* mm2 = x6 x4 x2 x0 */
219 
220  "movq 8(%0), %%mm5 \n\t" /* mm5 = x7 x5 x3 x1 */
221  "movq %%mm2, %%mm0 \n\t" /* mm0 = x6 x4 x2 x0 */
222 
223  "movq (%1), %%mm3 \n\t" /* mm3 = C6 C4 C2 C4 */
224  "movq %%mm5, %%mm6 \n\t" /* mm6 = x7 x5 x3 x1 */
225 
226  "punpckldq %%mm0, %%mm0 \n\t" /* mm0 = x2 x0 x2 x0 */
227 
228  "movq 8(%1), %%mm4 \n\t" /* mm4 = -C2 -C4 C6 C4 */
229  "pmaddwd %%mm0, %%mm3 \n\t" /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
230 
231  "movq 16(%1), %%mm1 \n\t" /* mm1 = -C7 C3 C3 C1 */
232  "punpckhdq %%mm2, %%mm2 \n\t" /* mm2 = x6 x4 x6 x4 */
233  :: "r" ((row+offset)), "r" (table)
234  );
235 }
236 
237 static inline void mmx_row (const int16_t * const table,
238  const int32_t * const rounder)
239 {
240  __asm__ volatile (
241  "pmaddwd %%mm2, %%mm4 \n\t" /* mm4 = -C4*x4-C2*x6 C4*x4+C6*x6 */
242  "punpckldq %%mm5, %%mm5 \n\t" /* mm5 = x3 x1 x3 x1 */
243 
244  "pmaddwd 32(%0), %%mm0 \n\t" /* mm0 = C4*x0-C2*x2 C4*x0-C6*x2 */
245  "punpckhdq %%mm6, %%mm6 \n\t" /* mm6 = x7 x5 x7 x5 */
246 
247  "movq 24(%0), %%mm7 \n\t" /* mm7 = -C5 -C1 C7 C5 */
248  "pmaddwd %%mm5, %%mm1 \n\t" /* mm1 = C3*x1-C7*x3 C1*x1+C3*x3 */
249 
250  "paddd (%1), %%mm3 \n\t" /* mm3 += rounder */
251  "pmaddwd %%mm6, %%mm7 \n\t" /* mm7 = -C1*x5-C5*x7 C5*x5+C7*x7 */
252 
253  "pmaddwd 40(%0), %%mm2 \n\t" /* mm2 = C4*x4-C6*x6 -C4*x4+C2*x6 */
254  "paddd %%mm4, %%mm3 \n\t" /* mm3 = a1 a0 + rounder */
255 
256  "pmaddwd 48(%0), %%mm5 \n\t" /* mm5 = C7*x1-C5*x3 C5*x1-C1*x3 */
257  "movq %%mm3, %%mm4 \n\t" /* mm4 = a1 a0 + rounder */
258 
259  "pmaddwd 56(%0), %%mm6 \n\t" /* mm6 = C3*x5-C1*x7 C7*x5+C3*x7 */
260  "paddd %%mm7, %%mm1 \n\t" /* mm1 = b1 b0 */
261 
262  "paddd (%1), %%mm0 \n\t" /* mm0 += rounder */
263  "psubd %%mm1, %%mm3 \n\t" /* mm3 = a1-b1 a0-b0 + rounder */
264 
265  "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm3 \n\t" /* mm3 = y6 y7 */
266  "paddd %%mm4, %%mm1 \n\t" /* mm1 = a1+b1 a0+b0 + rounder */
267 
268  "paddd %%mm2, %%mm0 \n\t" /* mm0 = a3 a2 + rounder */
269  "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm1 \n\t" /* mm1 = y1 y0 */
270 
271  "paddd %%mm6, %%mm5 \n\t" /* mm5 = b3 b2 */
272  "movq %%mm0, %%mm7 \n\t" /* mm7 = a3 a2 + rounder */
273 
274  "paddd %%mm5, %%mm0 \n\t" /* mm0 = a3+b3 a2+b2 + rounder */
275  "psubd %%mm5, %%mm7 \n\t" /* mm7 = a3-b3 a2-b2 + rounder */
276  :: "r" (table), "r" (rounder)
277  );
278 }
279 
280 static inline void mmx_row_tail (int16_t * const row, const int store)
281 {
282  __asm__ volatile (
283  "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t" /* mm0 = y3 y2 */
284 
285  "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm7 \n\t" /* mm7 = y4 y5 */
286 
287  "packssdw %%mm0, %%mm1 \n\t" /* mm1 = y3 y2 y1 y0 */
288 
289  "packssdw %%mm3, %%mm7 \n\t" /* mm7 = y6 y7 y4 y5 */
290 
291  "movq %%mm1, (%0) \n\t" /* save y3 y2 y1 y0 */
292  "movq %%mm7, %%mm4 \n\t" /* mm4 = y6 y7 y4 y5 */
293 
294  "pslld $16, %%mm7 \n\t" /* mm7 = y7 0 y5 0 */
295 
296  "psrld $16, %%mm4 \n\t" /* mm4 = 0 y6 0 y4 */
297 
298  "por %%mm4, %%mm7 \n\t" /* mm7 = y7 y6 y5 y4 */
299 
300  /* slot */
301 
302  "movq %%mm7, 8(%0) \n\t" /* save y7 y6 y5 y4 */
303  :: "r" (row+store)
304  );
305 }
306 
307 static inline void mmx_row_mid (int16_t * const row, const int store,
308  const int offset, const int16_t * const table)
309 {
310 
311  __asm__ volatile (
312  "movq (%0,%1), %%mm2 \n\t" /* mm2 = x6 x4 x2 x0 */
313  "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t" /* mm0 = y3 y2 */
314 
315  "movq 8(%0,%1), %%mm5 \n\t" /* mm5 = x7 x5 x3 x1 */
316  "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm7 \n\t" /* mm7 = y4 y5 */
317 
318  "packssdw %%mm0, %%mm1 \n\t" /* mm1 = y3 y2 y1 y0 */
319  "movq %%mm5, %%mm6 \n\t" /* mm6 = x7 x5 x3 x1 */
320 
321  "packssdw %%mm3, %%mm7 \n\t" /* mm7 = y6 y7 y4 y5 */
322  "movq %%mm2, %%mm0 \n\t" /* mm0 = x6 x4 x2 x0 */
323 
324  "movq %%mm1, (%0,%2) \n\t" /* save y3 y2 y1 y0 */
325  "movq %%mm7, %%mm1 \n\t" /* mm1 = y6 y7 y4 y5 */
326 
327  "punpckldq %%mm0, %%mm0 \n\t" /* mm0 = x2 x0 x2 x0 */
328  "psrld $16, %%mm7 \n\t" /* mm7 = 0 y6 0 y4 */
329 
330  "movq (%3), %%mm3 \n\t" /* mm3 = C6 C4 C2 C4 */
331  "pslld $16, %%mm1 \n\t" /* mm1 = y7 0 y5 0 */
332 
333  "movq 8(%3), %%mm4 \n\t" /* mm4 = -C2 -C4 C6 C4 */
334  "por %%mm1, %%mm7 \n\t" /* mm7 = y7 y6 y5 y4 */
335 
336  "movq 16(%3), %%mm1 \n\t" /* mm1 = -C7 C3 C3 C1 */
337  "punpckhdq %%mm2, %%mm2 \n\t" /* mm2 = x6 x4 x6 x4 */
338 
339  "movq %%mm7, 8(%0,%2) \n\t" /* save y7 y6 y5 y4 */
340  "pmaddwd %%mm0, %%mm3 \n\t" /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
341  : : "r" (row), "r" ((x86_reg) (2*offset)), "r" ((x86_reg) (2*store)), "r" (table)
342  );
343 }
344 
345 
346 #if 0
347 /* C column IDCT - it is just here to document the MMXEXT and MMX versions */
348 static inline void idct_col (int16_t * col, int offset)
349 {
350 /* multiplication - as implemented on mmx */
351 #define F(c,x) (((c) * (x)) >> 16)
352 
353 /* saturation - it helps us handle torture test cases */
354 #define S(x) (((x)>32767) ? 32767 : ((x)<-32768) ? -32768 : (x))
355 
356  int16_t x0, x1, x2, x3, x4, x5, x6, x7;
357  int16_t y0, y1, y2, y3, y4, y5, y6, y7;
358  int16_t a0, a1, a2, a3, b0, b1, b2, b3;
359  int16_t u04, v04, u26, v26, u17, v17, u35, v35, u12, v12;
360 
361  col += offset;
362 
363  x0 = col[0*8];
364  x1 = col[1*8];
365  x2 = col[2*8];
366  x3 = col[3*8];
367  x4 = col[4*8];
368  x5 = col[5*8];
369  x6 = col[6*8];
370  x7 = col[7*8];
371 
372  u04 = S (x0 + x4);
373  v04 = S (x0 - x4);
374  u26 = S (F (T2, x6) + x2);
375  v26 = S (F (T2, x2) - x6);
376 
377  a0 = S (u04 + u26);
378  a1 = S (v04 + v26);
379  a2 = S (v04 - v26);
380  a3 = S (u04 - u26);
381 
382  u17 = S (F (T1, x7) + x1);
383  v17 = S (F (T1, x1) - x7);
384  u35 = S (F (T3, x5) + x3);
385  v35 = S (F (T3, x3) - x5);
386 
387  b0 = S (u17 + u35);
388  b3 = S (v17 - v35);
389  u12 = S (u17 - u35);
390  v12 = S (v17 + v35);
391  u12 = S (2 * F (C4, u12));
392  v12 = S (2 * F (C4, v12));
393  b1 = S (u12 + v12);
394  b2 = S (u12 - v12);
395 
396  y0 = S (a0 + b0) >> COL_SHIFT;
397  y1 = S (a1 + b1) >> COL_SHIFT;
398  y2 = S (a2 + b2) >> COL_SHIFT;
399  y3 = S (a3 + b3) >> COL_SHIFT;
400 
401  y4 = S (a3 - b3) >> COL_SHIFT;
402  y5 = S (a2 - b2) >> COL_SHIFT;
403  y6 = S (a1 - b1) >> COL_SHIFT;
404  y7 = S (a0 - b0) >> COL_SHIFT;
405 
406  col[0*8] = y0;
407  col[1*8] = y1;
408  col[2*8] = y2;
409  col[3*8] = y3;
410  col[4*8] = y4;
411  col[5*8] = y5;
412  col[6*8] = y6;
413  col[7*8] = y7;
414 }
415 #endif
416 
417 
418 /* MMX column IDCT */
419 static inline void idct_col (int16_t * const col, const int offset)
420 {
421 #define T1 13036
422 #define T2 27146
423 #define T3 43790
424 #define C4 23170
425 
426  DECLARE_ALIGNED(8, static const short, t1_vector)[] = {
427  T1,T1,T1,T1,
428  T2,T2,T2,T2,
429  T3,T3,T3,T3,
430  C4,C4,C4,C4
431  };
432 
433  /* column code adapted from Peter Gubanov */
434  /* http://www.elecard.com/peter/idct.shtml */
435 
436  __asm__ volatile (
437  "movq (%0), %%mm0 \n\t" /* mm0 = T1 */
438 
439  "movq 2*8(%1), %%mm1 \n\t" /* mm1 = x1 */
440  "movq %%mm0, %%mm2 \n\t" /* mm2 = T1 */
441 
442  "movq 7*2*8(%1), %%mm4 \n\t" /* mm4 = x7 */
443  "pmulhw %%mm1, %%mm0 \n\t" /* mm0 = T1*x1 */
444 
445  "movq 16(%0), %%mm5 \n\t" /* mm5 = T3 */
446  "pmulhw %%mm4, %%mm2 \n\t" /* mm2 = T1*x7 */
447 
448  "movq 2*5*8(%1), %%mm6 \n\t" /* mm6 = x5 */
449  "movq %%mm5, %%mm7 \n\t" /* mm7 = T3-1 */
450 
451  "movq 3*8*2(%1), %%mm3 \n\t" /* mm3 = x3 */
452  "psubsw %%mm4, %%mm0 \n\t" /* mm0 = v17 */
453 
454  "movq 8(%0), %%mm4 \n\t" /* mm4 = T2 */
455  "pmulhw %%mm3, %%mm5 \n\t" /* mm5 = (T3-1)*x3 */
456 
457  "paddsw %%mm2, %%mm1 \n\t" /* mm1 = u17 */
458  "pmulhw %%mm6, %%mm7 \n\t" /* mm7 = (T3-1)*x5 */
459 
460  /* slot */
461 
462  "movq %%mm4, %%mm2 \n\t" /* mm2 = T2 */
463  "paddsw %%mm3, %%mm5 \n\t" /* mm5 = T3*x3 */
464 
465  "pmulhw 2*8*2(%1), %%mm4 \n\t" /* mm4 = T2*x2 */
466  "paddsw %%mm6, %%mm7 \n\t" /* mm7 = T3*x5 */
467 
468  "psubsw %%mm6, %%mm5 \n\t" /* mm5 = v35 */
469  "paddsw %%mm3, %%mm7 \n\t" /* mm7 = u35 */
470 
471  "movq 6*8*2(%1), %%mm3 \n\t" /* mm3 = x6 */
472  "movq %%mm0, %%mm6 \n\t" /* mm6 = v17 */
473 
474  "pmulhw %%mm3, %%mm2 \n\t" /* mm2 = T2*x6 */
475  "psubsw %%mm5, %%mm0 \n\t" /* mm0 = b3 */
476 
477  "psubsw %%mm3, %%mm4 \n\t" /* mm4 = v26 */
478  "paddsw %%mm6, %%mm5 \n\t" /* mm5 = v12 */
479 
480  "movq %%mm0, 3*8*2(%1)\n\t" /* save b3 in scratch0 */
481  "movq %%mm1, %%mm6 \n\t" /* mm6 = u17 */
482 
483  "paddsw 2*8*2(%1), %%mm2 \n\t" /* mm2 = u26 */
484  "paddsw %%mm7, %%mm6 \n\t" /* mm6 = b0 */
485 
486  "psubsw %%mm7, %%mm1 \n\t" /* mm1 = u12 */
487  "movq %%mm1, %%mm7 \n\t" /* mm7 = u12 */
488 
489  "movq 0*8(%1), %%mm3 \n\t" /* mm3 = x0 */
490  "paddsw %%mm5, %%mm1 \n\t" /* mm1 = u12+v12 */
491 
492  "movq 24(%0), %%mm0 \n\t" /* mm0 = C4/2 */
493  "psubsw %%mm5, %%mm7 \n\t" /* mm7 = u12-v12 */
494 
495  "movq %%mm6, 5*8*2(%1)\n\t" /* save b0 in scratch1 */
496  "pmulhw %%mm0, %%mm1 \n\t" /* mm1 = b1/2 */
497 
498  "movq %%mm4, %%mm6 \n\t" /* mm6 = v26 */
499  "pmulhw %%mm0, %%mm7 \n\t" /* mm7 = b2/2 */
500 
501  "movq 4*8*2(%1), %%mm5 \n\t" /* mm5 = x4 */
502  "movq %%mm3, %%mm0 \n\t" /* mm0 = x0 */
503 
504  "psubsw %%mm5, %%mm3 \n\t" /* mm3 = v04 */
505  "paddsw %%mm5, %%mm0 \n\t" /* mm0 = u04 */
506 
507  "paddsw %%mm3, %%mm4 \n\t" /* mm4 = a1 */
508  "movq %%mm0, %%mm5 \n\t" /* mm5 = u04 */
509 
510  "psubsw %%mm6, %%mm3 \n\t" /* mm3 = a2 */
511  "paddsw %%mm2, %%mm5 \n\t" /* mm5 = a0 */
512 
513  "paddsw %%mm1, %%mm1 \n\t" /* mm1 = b1 */
514  "psubsw %%mm2, %%mm0 \n\t" /* mm0 = a3 */
515 
516  "paddsw %%mm7, %%mm7 \n\t" /* mm7 = b2 */
517  "movq %%mm3, %%mm2 \n\t" /* mm2 = a2 */
518 
519  "movq %%mm4, %%mm6 \n\t" /* mm6 = a1 */
520  "paddsw %%mm7, %%mm3 \n\t" /* mm3 = a2+b2 */
521 
522  "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm3\n\t" /* mm3 = y2 */
523  "paddsw %%mm1, %%mm4\n\t" /* mm4 = a1+b1 */
524 
525  "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm4\n\t" /* mm4 = y1 */
526  "psubsw %%mm1, %%mm6 \n\t" /* mm6 = a1-b1 */
527 
528  "movq 5*8*2(%1), %%mm1 \n\t" /* mm1 = b0 */
529  "psubsw %%mm7, %%mm2 \n\t" /* mm2 = a2-b2 */
530 
531  "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm6\n\t" /* mm6 = y6 */
532  "movq %%mm5, %%mm7 \n\t" /* mm7 = a0 */
533 
534  "movq %%mm4, 1*8*2(%1)\n\t" /* save y1 */
535  "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm2\n\t" /* mm2 = y5 */
536 
537  "movq %%mm3, 2*8*2(%1)\n\t" /* save y2 */
538  "paddsw %%mm1, %%mm5 \n\t" /* mm5 = a0+b0 */
539 
540  "movq 3*8*2(%1), %%mm4 \n\t" /* mm4 = b3 */
541  "psubsw %%mm1, %%mm7 \n\t" /* mm7 = a0-b0 */
542 
543  "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm5\n\t" /* mm5 = y0 */
544  "movq %%mm0, %%mm3 \n\t" /* mm3 = a3 */
545 
546  "movq %%mm2, 5*8*2(%1)\n\t" /* save y5 */
547  "psubsw %%mm4, %%mm3 \n\t" /* mm3 = a3-b3 */
548 
549  "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm7\n\t" /* mm7 = y7 */
550  "paddsw %%mm0, %%mm4 \n\t" /* mm4 = a3+b3 */
551 
552  "movq %%mm5, 0*8*2(%1)\n\t" /* save y0 */
553  "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm3\n\t" /* mm3 = y4 */
554 
555  "movq %%mm6, 6*8*2(%1)\n\t" /* save y6 */
556  "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm4\n\t" /* mm4 = y3 */
557 
558  "movq %%mm7, 7*8*2(%1)\n\t" /* save y7 */
559 
560  "movq %%mm3, 4*8*2(%1)\n\t" /* save y4 */
561 
562  "movq %%mm4, 3*8*2(%1)\n\t" /* save y3 */
563  :: "r" (t1_vector), "r" (col+offset)
564  );
565 
566 #undef T1
567 #undef T2
568 #undef T3
569 #undef C4
570 }
571 
572 
573 DECLARE_ALIGNED(8, static const int32_t, rounder0)[] =
574  rounder ((1 << (COL_SHIFT - 1)) - 0.5);
575 DECLARE_ALIGNED(8, static const int32_t, rounder4)[] = rounder (0);
576 DECLARE_ALIGNED(8, static const int32_t, rounder1)[] =
577  rounder (1.25683487303); /* C1*(C1/C4+C1+C7)/2 */
578 DECLARE_ALIGNED(8, static const int32_t, rounder7)[] =
579  rounder (-0.25); /* C1*(C7/C4+C7-C1)/2 */
580 DECLARE_ALIGNED(8, static const int32_t, rounder2)[] =
581  rounder (0.60355339059); /* C2 * (C6+C2)/2 */
582 DECLARE_ALIGNED(8, static const int32_t, rounder6)[] =
583  rounder (-0.25); /* C2 * (C6-C2)/2 */
584 DECLARE_ALIGNED(8, static const int32_t, rounder3)[] =
585  rounder (0.087788325588); /* C3*(-C3/C4+C3+C5)/2 */
586 DECLARE_ALIGNED(8, static const int32_t, rounder5)[] =
587  rounder (-0.441341716183); /* C3*(-C5/C4+C5-C3)/2 */
588 
589 #undef COL_SHIFT
590 #undef ROW_SHIFT
591 
592 #define declare_idct(idct,table,idct_row_head,idct_row,idct_row_tail,idct_row_mid) \
593 void idct (int16_t * const block) \
594 { \
595  DECLARE_ALIGNED(16, static const int16_t, table04)[] = \
596  table (22725, 21407, 19266, 16384, 12873, 8867, 4520); \
597  DECLARE_ALIGNED(16, static const int16_t, table17)[] = \
598  table (31521, 29692, 26722, 22725, 17855, 12299, 6270); \
599  DECLARE_ALIGNED(16, static const int16_t, table26)[] = \
600  table (29692, 27969, 25172, 21407, 16819, 11585, 5906); \
601  DECLARE_ALIGNED(16, static const int16_t, table35)[] = \
602  table (26722, 25172, 22654, 19266, 15137, 10426, 5315); \
603  \
604  idct_row_head (block, 0*8, table04); \
605  idct_row (table04, rounder0); \
606  idct_row_mid (block, 0*8, 4*8, table04); \
607  idct_row (table04, rounder4); \
608  idct_row_mid (block, 4*8, 1*8, table17); \
609  idct_row (table17, rounder1); \
610  idct_row_mid (block, 1*8, 7*8, table17); \
611  idct_row (table17, rounder7); \
612  idct_row_mid (block, 7*8, 2*8, table26); \
613  idct_row (table26, rounder2); \
614  idct_row_mid (block, 2*8, 6*8, table26); \
615  idct_row (table26, rounder6); \
616  idct_row_mid (block, 6*8, 3*8, table35); \
617  idct_row (table35, rounder3); \
618  idct_row_mid (block, 3*8, 5*8, table35); \
619  idct_row (table35, rounder5); \
620  idct_row_tail (block, 5*8); \
621  \
622  idct_col (block, 0); \
623  idct_col (block, 4); \
624 }
625 
626 declare_idct (ff_mmxext_idct, mmxext_table,
627  mmxext_row_head, mmxext_row, mmxext_row_tail, mmxext_row_mid)
628 
629 declare_idct (ff_mmx_idct, mmx_table,
630  mmx_row_head, mmx_row, mmx_row_tail, mmx_row_mid)
631 
632 #endif /* HAVE_INLINE_ASM */