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/asm.h"
26 #include "dsputil_mmx.h"
33 #define round(bias) ((int)(((bias)+0.5) * (1<<ROW_SHIFT)))
34 #define rounder(bias) {round (bias), round (bias)}
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)
42 int C1, C2, C3, C4, C5, C6, C7;
43 int a0, a1, a2, a3, b0, b1, b2, b3;
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;
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];
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;
79 #define mmxext_table(c1,c2,c3,c4,c5,c6,c7) { c4, c2, -c4, -c2, \
88 static inline void mmxext_row_head (int16_t * const row, const int offset,
89 const int16_t * const table)
92 "movq (%0), %%mm2 \n\t" /* mm2 = x6 x4 x2 x0 */
94 "movq 8(%0), %%mm5 \n\t" /* mm5 = x7 x5 x3 x1 */
95 "movq %%mm2, %%mm0 \n\t" /* mm0 = x6 x4 x2 x0 */
97 "movq (%1), %%mm3 \n\t" /* mm3 = -C2 -C4 C2 C4 */
98 "movq %%mm5, %%mm6 \n\t" /* mm6 = x7 x5 x3 x1 */
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 */
103 "pshufw $0x4e, %%mm2, %%mm2 \n\t" /* mm2 = x2 x0 x6 x4 */
104 :: "r" ((row+offset)), "r" (table)
108 static inline void mmxext_row (const int16_t * const table,
109 const int32_t * const rounder)
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 */
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 */
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 */
121 "paddd (%1), %%mm3 \n\t" /* mm3 += rounder */
122 "pmaddwd %%mm6, %%mm7 \n\t" /* mm7 = C3*x1-C7*x3 C5*x5+C7*x7 */
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 */
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 */
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 */
133 "paddd (%1), %%mm0 \n\t" /* mm0 += rounder */
134 "psubd %%mm1, %%mm3 \n\t" /* mm3 = a1-b1 a0-b0 + rounder */
136 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm3 \n\t" /* mm3 = y6 y7 */
137 "paddd %%mm4, %%mm1 \n\t" /* mm1 = a1+b1 a0+b0 + rounder */
139 "paddd %%mm2, %%mm0 \n\t" /* mm0 = a3 a2 + rounder */
140 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm1 \n\t" /* mm1 = y1 y0 */
142 "paddd %%mm6, %%mm5 \n\t" /* mm5 = b3 b2 */
143 "movq %%mm0, %%mm4 \n\t" /* mm4 = a3 a2 + rounder */
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));
150 static inline void mmxext_row_tail (int16_t * const row, const int store)
153 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t" /* mm0 = y3 y2 */
155 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm4 \n\t" /* mm4 = y4 y5 */
157 "packssdw %%mm0, %%mm1 \n\t" /* mm1 = y3 y2 y1 y0 */
159 "packssdw %%mm3, %%mm4 \n\t" /* mm4 = y6 y7 y4 y5 */
161 "movq %%mm1, (%0) \n\t" /* save y3 y2 y1 y0 */
162 "pshufw $0xb1, %%mm4, %%mm4 \n\t" /* mm4 = y7 y6 y5 y4 */
166 "movq %%mm4, 8(%0) \n\t" /* save y7 y6 y5 y4 */
171 static inline void mmxext_row_mid (int16_t * const row, const int store,
173 const int16_t * const table)
176 "movq (%0,%1), %%mm2 \n\t" /* mm2 = x6 x4 x2 x0 */
177 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t" /* mm0 = y3 y2 */
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 */
182 "packssdw %%mm0, %%mm1 \n\t" /* mm1 = y3 y2 y1 y0 */
183 "movq %%mm5, %%mm6 \n\t" /* mm6 = x7 x5 x3 x1 */
185 "packssdw %%mm3, %%mm4 \n\t" /* mm4 = y6 y7 y4 y5 */
186 "movq %%mm2, %%mm0 \n\t" /* mm0 = x6 x4 x2 x0 */
188 "movq %%mm1, (%0,%2) \n\t" /* save y3 y2 y1 y0 */
189 "pshufw $0xb1, %%mm4, %%mm4\n\t" /* mm4 = y7 y6 y5 y4 */
191 "movq (%3), %%mm3 \n\t" /* mm3 = -C2 -C4 C2 C4 */
192 "movq %%mm4, 8(%0,%2) \n\t" /* save y7 y6 y5 y4 */
194 "pmaddwd %%mm0, %%mm3 \n\t" /* mm3= -C4*x4-C2*x6 C4*x0+C2*x2 */
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)
205 #define mmx_table(c1,c2,c3,c4,c5,c6,c7) { c4, c2, c4, c6, \
214 static inline void mmx_row_head (int16_t * const row, const int offset,
215 const int16_t * const table)
218 "movq (%0), %%mm2 \n\t" /* mm2 = x6 x4 x2 x0 */
220 "movq 8(%0), %%mm5 \n\t" /* mm5 = x7 x5 x3 x1 */
221 "movq %%mm2, %%mm0 \n\t" /* mm0 = x6 x4 x2 x0 */
223 "movq (%1), %%mm3 \n\t" /* mm3 = C6 C4 C2 C4 */
224 "movq %%mm5, %%mm6 \n\t" /* mm6 = x7 x5 x3 x1 */
226 "punpckldq %%mm0, %%mm0 \n\t" /* mm0 = x2 x0 x2 x0 */
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 */
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)
237 static inline void mmx_row (const int16_t * const table,
238 const int32_t * const rounder)
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 */
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 */
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 */
250 "paddd (%1), %%mm3 \n\t" /* mm3 += rounder */
251 "pmaddwd %%mm6, %%mm7 \n\t" /* mm7 = -C1*x5-C5*x7 C5*x5+C7*x7 */
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 */
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 */
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 */
262 "paddd (%1), %%mm0 \n\t" /* mm0 += rounder */
263 "psubd %%mm1, %%mm3 \n\t" /* mm3 = a1-b1 a0-b0 + rounder */
265 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm3 \n\t" /* mm3 = y6 y7 */
266 "paddd %%mm4, %%mm1 \n\t" /* mm1 = a1+b1 a0+b0 + rounder */
268 "paddd %%mm2, %%mm0 \n\t" /* mm0 = a3 a2 + rounder */
269 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm1 \n\t" /* mm1 = y1 y0 */
271 "paddd %%mm6, %%mm5 \n\t" /* mm5 = b3 b2 */
272 "movq %%mm0, %%mm7 \n\t" /* mm7 = a3 a2 + rounder */
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)
280 static inline void mmx_row_tail (int16_t * const row, const int store)
283 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t" /* mm0 = y3 y2 */
285 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm7 \n\t" /* mm7 = y4 y5 */
287 "packssdw %%mm0, %%mm1 \n\t" /* mm1 = y3 y2 y1 y0 */
289 "packssdw %%mm3, %%mm7 \n\t" /* mm7 = y6 y7 y4 y5 */
291 "movq %%mm1, (%0) \n\t" /* save y3 y2 y1 y0 */
292 "movq %%mm7, %%mm4 \n\t" /* mm4 = y6 y7 y4 y5 */
294 "pslld $16, %%mm7 \n\t" /* mm7 = y7 0 y5 0 */
296 "psrld $16, %%mm4 \n\t" /* mm4 = 0 y6 0 y4 */
298 "por %%mm4, %%mm7 \n\t" /* mm7 = y7 y6 y5 y4 */
302 "movq %%mm7, 8(%0) \n\t" /* save y7 y6 y5 y4 */
307 static inline void mmx_row_mid (int16_t * const row, const int store,
308 const int offset, const int16_t * const table)
312 "movq (%0,%1), %%mm2 \n\t" /* mm2 = x6 x4 x2 x0 */
313 "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t" /* mm0 = y3 y2 */
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 */
318 "packssdw %%mm0, %%mm1 \n\t" /* mm1 = y3 y2 y1 y0 */
319 "movq %%mm5, %%mm6 \n\t" /* mm6 = x7 x5 x3 x1 */
321 "packssdw %%mm3, %%mm7 \n\t" /* mm7 = y6 y7 y4 y5 */
322 "movq %%mm2, %%mm0 \n\t" /* mm0 = x6 x4 x2 x0 */
324 "movq %%mm1, (%0,%2) \n\t" /* save y3 y2 y1 y0 */
325 "movq %%mm7, %%mm1 \n\t" /* mm1 = y6 y7 y4 y5 */
327 "punpckldq %%mm0, %%mm0 \n\t" /* mm0 = x2 x0 x2 x0 */
328 "psrld $16, %%mm7 \n\t" /* mm7 = 0 y6 0 y4 */
330 "movq (%3), %%mm3 \n\t" /* mm3 = C6 C4 C2 C4 */
331 "pslld $16, %%mm1 \n\t" /* mm1 = y7 0 y5 0 */
333 "movq 8(%3), %%mm4 \n\t" /* mm4 = -C2 -C4 C6 C4 */
334 "por %%mm1, %%mm7 \n\t" /* mm7 = y7 y6 y5 y4 */
336 "movq 16(%3), %%mm1 \n\t" /* mm1 = -C7 C3 C3 C1 */
337 "punpckhdq %%mm2, %%mm2 \n\t" /* mm2 = x6 x4 x6 x4 */
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)
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)
350 /* multiplication - as implemented on mmx */
351 #define F(c,x) (((c) * (x)) >> 16)
353 /* saturation - it helps us handle torture test cases */
354 #define S(x) (((x)>32767) ? 32767 : ((x)<-32768) ? -32768 : (x))
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;
374 u26 = S (F (T2, x6) + x2);
375 v26 = S (F (T2, x2) - x6);
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);
391 u12 = S (2 * F (C4, u12));
392 v12 = S (2 * F (C4, v12));
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;
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;
418 /* MMX column IDCT */
419 static inline void idct_col (int16_t * const col, const int offset)
426 DECLARE_ALIGNED(8, static const short, t1_vector)[] = {
433 /* column code adapted from Peter Gubanov */
434 /* http://www.elecard.com/peter/idct.shtml */
437 "movq (%0), %%mm0 \n\t" /* mm0 = T1 */
439 "movq 2*8(%1), %%mm1 \n\t" /* mm1 = x1 */
440 "movq %%mm0, %%mm2 \n\t" /* mm2 = T1 */
442 "movq 7*2*8(%1), %%mm4 \n\t" /* mm4 = x7 */
443 "pmulhw %%mm1, %%mm0 \n\t" /* mm0 = T1*x1 */
445 "movq 16(%0), %%mm5 \n\t" /* mm5 = T3 */
446 "pmulhw %%mm4, %%mm2 \n\t" /* mm2 = T1*x7 */
448 "movq 2*5*8(%1), %%mm6 \n\t" /* mm6 = x5 */
449 "movq %%mm5, %%mm7 \n\t" /* mm7 = T3-1 */
451 "movq 3*8*2(%1), %%mm3 \n\t" /* mm3 = x3 */
452 "psubsw %%mm4, %%mm0 \n\t" /* mm0 = v17 */
454 "movq 8(%0), %%mm4 \n\t" /* mm4 = T2 */
455 "pmulhw %%mm3, %%mm5 \n\t" /* mm5 = (T3-1)*x3 */
457 "paddsw %%mm2, %%mm1 \n\t" /* mm1 = u17 */
458 "pmulhw %%mm6, %%mm7 \n\t" /* mm7 = (T3-1)*x5 */
462 "movq %%mm4, %%mm2 \n\t" /* mm2 = T2 */
463 "paddsw %%mm3, %%mm5 \n\t" /* mm5 = T3*x3 */
465 "pmulhw 2*8*2(%1), %%mm4 \n\t" /* mm4 = T2*x2 */
466 "paddsw %%mm6, %%mm7 \n\t" /* mm7 = T3*x5 */
468 "psubsw %%mm6, %%mm5 \n\t" /* mm5 = v35 */
469 "paddsw %%mm3, %%mm7 \n\t" /* mm7 = u35 */
471 "movq 6*8*2(%1), %%mm3 \n\t" /* mm3 = x6 */
472 "movq %%mm0, %%mm6 \n\t" /* mm6 = v17 */
474 "pmulhw %%mm3, %%mm2 \n\t" /* mm2 = T2*x6 */
475 "psubsw %%mm5, %%mm0 \n\t" /* mm0 = b3 */
477 "psubsw %%mm3, %%mm4 \n\t" /* mm4 = v26 */
478 "paddsw %%mm6, %%mm5 \n\t" /* mm5 = v12 */
480 "movq %%mm0, 3*8*2(%1)\n\t" /* save b3 in scratch0 */
481 "movq %%mm1, %%mm6 \n\t" /* mm6 = u17 */
483 "paddsw 2*8*2(%1), %%mm2 \n\t" /* mm2 = u26 */
484 "paddsw %%mm7, %%mm6 \n\t" /* mm6 = b0 */
486 "psubsw %%mm7, %%mm1 \n\t" /* mm1 = u12 */
487 "movq %%mm1, %%mm7 \n\t" /* mm7 = u12 */
489 "movq 0*8(%1), %%mm3 \n\t" /* mm3 = x0 */
490 "paddsw %%mm5, %%mm1 \n\t" /* mm1 = u12+v12 */
492 "movq 24(%0), %%mm0 \n\t" /* mm0 = C4/2 */
493 "psubsw %%mm5, %%mm7 \n\t" /* mm7 = u12-v12 */
495 "movq %%mm6, 5*8*2(%1)\n\t" /* save b0 in scratch1 */
496 "pmulhw %%mm0, %%mm1 \n\t" /* mm1 = b1/2 */
498 "movq %%mm4, %%mm6 \n\t" /* mm6 = v26 */
499 "pmulhw %%mm0, %%mm7 \n\t" /* mm7 = b2/2 */
501 "movq 4*8*2(%1), %%mm5 \n\t" /* mm5 = x4 */
502 "movq %%mm3, %%mm0 \n\t" /* mm0 = x0 */
504 "psubsw %%mm5, %%mm3 \n\t" /* mm3 = v04 */
505 "paddsw %%mm5, %%mm0 \n\t" /* mm0 = u04 */
507 "paddsw %%mm3, %%mm4 \n\t" /* mm4 = a1 */
508 "movq %%mm0, %%mm5 \n\t" /* mm5 = u04 */
510 "psubsw %%mm6, %%mm3 \n\t" /* mm3 = a2 */
511 "paddsw %%mm2, %%mm5 \n\t" /* mm5 = a0 */
513 "paddsw %%mm1, %%mm1 \n\t" /* mm1 = b1 */
514 "psubsw %%mm2, %%mm0 \n\t" /* mm0 = a3 */
516 "paddsw %%mm7, %%mm7 \n\t" /* mm7 = b2 */
517 "movq %%mm3, %%mm2 \n\t" /* mm2 = a2 */
519 "movq %%mm4, %%mm6 \n\t" /* mm6 = a1 */
520 "paddsw %%mm7, %%mm3 \n\t" /* mm3 = a2+b2 */
522 "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm3\n\t" /* mm3 = y2 */
523 "paddsw %%mm1, %%mm4\n\t" /* mm4 = a1+b1 */
525 "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm4\n\t" /* mm4 = y1 */
526 "psubsw %%mm1, %%mm6 \n\t" /* mm6 = a1-b1 */
528 "movq 5*8*2(%1), %%mm1 \n\t" /* mm1 = b0 */
529 "psubsw %%mm7, %%mm2 \n\t" /* mm2 = a2-b2 */
531 "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm6\n\t" /* mm6 = y6 */
532 "movq %%mm5, %%mm7 \n\t" /* mm7 = a0 */
534 "movq %%mm4, 1*8*2(%1)\n\t" /* save y1 */
535 "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm2\n\t" /* mm2 = y5 */
537 "movq %%mm3, 2*8*2(%1)\n\t" /* save y2 */
538 "paddsw %%mm1, %%mm5 \n\t" /* mm5 = a0+b0 */
540 "movq 3*8*2(%1), %%mm4 \n\t" /* mm4 = b3 */
541 "psubsw %%mm1, %%mm7 \n\t" /* mm7 = a0-b0 */
543 "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm5\n\t" /* mm5 = y0 */
544 "movq %%mm0, %%mm3 \n\t" /* mm3 = a3 */
546 "movq %%mm2, 5*8*2(%1)\n\t" /* save y5 */
547 "psubsw %%mm4, %%mm3 \n\t" /* mm3 = a3-b3 */
549 "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm7\n\t" /* mm7 = y7 */
550 "paddsw %%mm0, %%mm4 \n\t" /* mm4 = a3+b3 */
552 "movq %%mm5, 0*8*2(%1)\n\t" /* save y0 */
553 "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm3\n\t" /* mm3 = y4 */
555 "movq %%mm6, 6*8*2(%1)\n\t" /* save y6 */
556 "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm4\n\t" /* mm4 = y3 */
558 "movq %%mm7, 7*8*2(%1)\n\t" /* save y7 */
560 "movq %%mm3, 4*8*2(%1)\n\t" /* save y4 */
562 "movq %%mm4, 3*8*2(%1)\n\t" /* save y3 */
563 :: "r" (t1_vector), "r" (col+offset)
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 */
592 #define declare_idct(idct,table,idct_row_head,idct_row,idct_row_tail,idct_row_mid) \
593 void idct (int16_t * const block) \
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); \
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); \
622 idct_col (block, 0); \
623 idct_col (block, 4); \
626 declare_idct (ff_mmxext_idct, mmxext_table,
627 mmxext_row_head, mmxext_row, mmxext_row_tail, mmxext_row_mid)
629 declare_idct (ff_mmx_idct, mmx_table,
630 mmx_row_head, mmx_row, mmx_row_tail, mmx_row_mid)
632 #endif /* HAVE_INLINE_ASM */